v11.7.0

Headline

• A new POST /v1/models/register endpoint registers or updates user.* model definitions without downloading any files.
• New GET/POST/DELETE /v1/models/{id}/options endpoints save, inspect, and reset per-model recipe options without loading the model, with per-load transient overrides.
• New GET /v1/stats and Prometheus /metrics endpoints report prefix-cache effectiveness and router route-switch counters.
• The built-in catalog adds the Qwen3.8-27B and NVIDIA Nemotron 3.5 Lightning 30B-A3B GGUF models and the Z-Image-Turbo image model.
• Lemonade can now be installed on Windows with winget and on macOS with a Homebrew cask.

Breaking Changes

• POST /pull now returns 400 and registers nothing when a model definition names an unservable or contradictory deployment mode; such labels are no longer silently normalized.
• Stored user_models.json entries that violate the new deployment-mode label rule are now skipped at startup with an error instead of being repaired.
• Removed the deprecated Lite Collection and Ultra Collection model registry entries.
• /v1/audio/transcriptions now returns raw text bodies for text/srt/vtt instead of a JSON {text: ...} wrapper, and returns 400 for unsupported response_format values.
• The TheNoise image backend no longer accepts per-request upscale and lora_dir recipe options; lora_dir is now set server-wide in config.json and upscale is only honored as a passed-through request parameter.
• lemonade launch opencode now emits a limit object (limit.context/limit.output) instead of a top-level contextWindow field.
• The no_broadcast config key is replaced by an inverted broadcast boolean, auto-migrated on load.
• An existing but unreadable extra_models_dir now returns 400 on /internal/set instead of being silently applied.

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_11.7.0-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_11.7.0-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-11.7.0-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-11.7.0-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-11.7.0-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-11.7.0-fc44.aarch64.rpm
macOS	Lemonade-11.7.0-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-11.7.0-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-11.7.0-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-11.7.0-windows-x64.zip
macOS arm64	lemonade-embeddable-11.7.0-macos-arm64.tar.gz

---

What's Changed

Thanks @SlawomirNowaczyk, @abn, @anditherobot, @bitgamma, @blackdeathdrow, @bong-water-water-bong, @fl0rianr, @ianbmacdonald, @iswaryaalex, @jeremyfowers, @kenvandine, @mzyy94, @osimarr, @popey, @ramkrishna2910, @superm1, @yelliver, @zaneni6 for your awesome contributions to this release!

Click to expand changelog

• Bump TheNoise to version v0.1.2 by @bitgamma in #3079
• docs(backends): drop the model catalog tables from the backend reference by @jeremyfowers in #3103
• fix(alias): report the real reason for alias validation failures by @ramkrishna2910 in #3111
• feat(server): add model registration endpoint by @fl0rianr in #3118
• Debian packaging improvements by @superm1 in #3039
• docs: add winget and Homebrew install to homepage quickstart by @yelliver in #2970
• Trigger snap release candidate build from release branches by @kenvandine in #2908
• fix(server): compare artifacts, not commit SHAs, when checking for model updates by @blackdeathdrow in #3073
• fix(server): normalize thinking controls for llama.cpp by @fl0rianr in #3132
• fix: stop request during prefill now possible by @fl0rianr in #3133
• ci: stop building unused targets in the validate workflows by @jeremyfowers in #3052
• fix(server): harden extra models directory handling by @fl0rianr in #3105
• Fix #1833: Improve transcription response format compatibility by @anditherobot in #2088
• ci: stop pinning lite validation legs to the 128gb runner by @jeremyfowers in #3139
• feat(models): add an explicit chat label by @jeremyfowers in #3099
• feat(telemetry): capture prefix-cache effectiveness (cache_n, cached_tokens) and route-switch counts by @ramkrishna2910 in #2968
• Update FLM version by @zaneni6 in #3144
• Backend Battle Nightly Regression by @iswaryaalex in #3069
• Remove emoji from FAQ title by @jeremyfowers in #3154
• Manage per-model recipe options without loading by @jeremyfowers in #3110
• Enable % used context on opencode side panel by @osimarr in #1725
• feat(server,cli): add discovery and broadcast controls with config decoupling by @abn in #3135
• fix(packaging): add AppStream MetaInfo files for desktop entries by @superm1 in #3123
• fix(flm): correctly report downloaded model size by @mzyy94 in #3166
• Minor refactor of new gate code by @SlawomirNowaczyk in #2971
• update TheNoise v0.2.1 by @bitgamma in #3171
• fix(flm): stop re-pulling already-downloaded models on every load by @mzyy94 in #3167
• feat: flag comment slop in a pre-commit rule, and run pre-commit in CI by @ianbmacdonald in #2689
• add lemonade version to benchmark results by @bitgamma in #3173
• feat(server): support transient null overrides on load by @fl0rianr in #3172
• fix(macos): reduce log directory permissions to prevent privilege esc… by @superm1 in #2625
• docs: add pull request template by @fl0rianr in #2551
• fix(npu): pre-reject NPU loads when auto-tune can only reserve fallback ctx (#1151) by @bong-water-water-bong in #3164
• models: add Qwen3.8 and Nemotron 3.5 GGUF by @yelliver in #3177
• Fix pre-commit JSON validation issues by @superm1 in #3179
• docs: fix multi-file checkpoint rendering by @yelliver in #3174
• Remove some MTP default flags by @bitgamma in #3187
• ci: wire server eviction tests into CI by @fl0rianr in #3015
• ci: run watchdog lifecycle tests in backend CI by @fl0rianr in #3021
• test: replace Python system info replicas with C++ coverage by @fl0rianr in #3020
• fix(server): let load args override saved args by @fl0rianr in https://github.com/lemonade-sdk/...

v11.6.0

Headline

• The Muse-Glimmer-30B-GGUF model joins the built-in llama.cpp catalog with draft decoding, vision, and tool-calling support.
• The llama.cpp ROCm backend now runs on AMD Instinct MI100, MI200, MI210, and MI250 GPUs on Linux.
• A new experimental TheNoise ROCm image-generation backend brings Anima and Krea-2 to AMD Strix Halo and Strix Point iGPUs.
• The new lemonade alias commands let you assign your own names to models, with alias-aware /v1/models listing and admin endpoints.
• lemonade bench adds opt-in vision benchmark scenarios for multimodal models.

Breaking Changes

• /v1/audio/speech with stream_format=audio and a non-PCM response_format (such as mp3) against Kokoro now returns HTTP 400 (only pcm is supported) instead of silently returning PCM audio.
• /v1/models now lists each GGUF quant in a split extra_models_dir subfolder as its own entry rather than one entry per folder; the old folder name still resolves as a hidden request alias.
• The Windows MSI layout changed: Lemonade_Server_MSI now contains only lemonade-server-minimal.msi, and the desktop lemonade.msi now ships in a new Lemonade_Desktop_MSI artifact, so automation that reads lemonade.msi from the old artifact must be repointed.

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_11.6.0-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_11.6.0-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-11.6.0-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-11.6.0-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-11.6.0-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-11.6.0-fc44.aarch64.rpm
macOS	Lemonade-11.6.0-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-11.6.0-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-11.6.0-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-11.6.0-windows-x64.zip
macOS arm64	lemonade-embeddable-11.6.0-macos-arm64.tar.gz

---

What's Changed

Thanks @Bekhouche, @SlawomirNowaczyk, @Yigtwxx, @ZMXJJ, @abn, @anditherobot, @bitgamma, @blackdeathdrow, @ckuethe, @duggiefresh, @fl0rianr, @github-actions, @hogeheer499-commits, @jeremyfowers, @kenvandine, @popey, @ramkrishna2910, @sreeram-11, @storm1er, @superm1 for your awesome contributions to this release!

Click to expand changelog

• fix(ci): prevent MSI verify flake from stale ARP entries on shared runners by @jeremyfowers in #2895
• docs: add testing guide by @jeremyfowers in #2891
• fix(build): patch libwebsockets GENHDR OUTPUT to silence MSB8065 by @blackdeathdrow in #2893
• ci: stop re-running the full suite on push to main by @jeremyfowers in #2899
• fix(cli): respect explicit --port when --host includes a scheme by @kenvandine in #2914
• README.md - Fix download links by @duggiefresh in #2827
• ci: run packaging, PPA, distro, macOS and backend-validation jobs in the merge queue by @jeremyfowers in #2911
• New helper combining add_test() and register_cpp_ci_test() by @SlawomirNowaczyk in #2877
• [llamacpp] replace deprecated flags, better IO flag handling by @bitgamma in #2833
• Add /rerank as the reranking endpoint path (alias /reranking) by @sreeram-11 in #2924
• fix(server): keep interrupted variantless model downloads resumable by @fl0rianr in #2876
• test: cover embeddings missing model error by @fl0rianr in #2061
• fix(docs): Added group_add configuration for ROCm support in Docker. by @storm1er in #2857
• ci: run .exe/.deb inference suites in the merge queue by @jeremyfowers in #2950
• ci: isolate llama.cpp validation cleanup on self-hosted runners by @fl0rianr in #2915
• Reclaim routing helpers when a router collection's policy changes by @SlawomirNowaczyk in #2795
• [backends] Add support for image generation through TheNoise by @bitgamma in #2927
• fix(ci): refactor cleanup and improve logging llama.cpp validation by @fl0rianr in #2992
• fix(server): frame backend errors as SSE events on the streaming path by @Yigtwxx in #2975
• ci: cut the longest test jobs via parameter tuning by @jeremyfowers in #2953
• docs: add new engine logos to homepage engine ticker by @jeremyfowers in #2920
• Show all local model versions in one folder by @anditherobot in #2107
• test: run the committed model-type classifier test in CI by @Yigtwxx in #2976
• fix(server): return the router error status on rerank, slots and tokenize by @Yigtwxx in #2974
• ci(test): speed up server job suite by @fl0rianr in #3007
• Add complete vision benchmark support by @ckuethe in #2869
• ci: cut installer build time on the PR/merge-queue critical path by @jeremyfowers in #2989
• fix(server): report backend errors on the Anthropic messages bridge by @Yigtwxx in #3006
• docs: add Muse Glimmer 30B blog post by @jeremyfowers in #3033
• fix: add GITHUB_TOKEN / GH_TOKEN support to GitHub API requests by @blackdeathdrow in #2995
• ci: fail fast in merge queue matrices by @fl0rianr in #3014
• feat(rocm): enable AMD Instinct MI100 (gfx908) and MI200 (gfx90a) in llama.cpp ROCm by @kenvandine in #2092
• feat(router): report estimated cost on collection.router decisions by @Bekhouche in #2763
• Allow explicit system llama.cpp backend by @popey in #3016
• feat(server): add model alias system with /v1/models listing and /internal/aliases endpoints by @abn in #2818
• fix(server): sum sibling shards when computing on-disk GGUF size by @blackdeathdrow in #2973
• fix(tts): stop dropping response_format on the streaming path by @ZMXJJ in #3029
• fix: honor custom backend binary environment variables by @hogeheer499-commits in #3004
• Update llama.cpp to b10360 by @github-actions[bot] in #3053
• Add support for a default model source by @superm1 in #3036
• feat(llamacpp): auto-detect draft GGUF companions by @fl0rianr in #3051
• Bump project version from 11.5.2 to 11.6.0 by @ramkrishna2910 in #3085
• feat(models): add Muse Glimmer 30B to the model catalog by @ramkrishna2910 in #3090
• Update llama.cpp to b10375 by @g...

v11.5.2

Headline

• The standalone macOS/Linux lemonade-tray client can now connect to servers over HTTPS/TLS.
• Repeatable llama.cpp custom arguments such as --override-kv now preserve every occurrence instead of collapsing to the last value.
• lemonade bench output now includes a hardware section covering CPU, GPU, RAM, OS, and backends.
• The llama.cpp backend now returns a clear error when an explicit device prefix contradicts the resolved backend.
• The Windows MSI installer now restores a saved custom per-user install location during upgrades.

Breaking Changes

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_11.5.2-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_11.5.2-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-11.5.2-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-11.5.2-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-11.5.2-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-11.5.2-fc44.aarch64.rpm
macOS	Lemonade-11.5.2-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-11.5.2-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-11.5.2-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-11.5.2-windows-x64.zip
macOS arm64	lemonade-embeddable-11.5.2-macos-arm64.tar.gz

---

What's Changed

Thanks @abn, @bitgamma, @ckuethe, @fl0rianr, @github-actions, @jeremyfowers, @kenvandine, @meghsat, @sujikathir for your awesome contributions to this release!

Click to expand changelog

• fix(test): stabilize semantic router similarity thresholds by @fl0rianr in #2824
• ci: add and scope concurrency policies across workflows by @jeremyfowers in #2851
• feat(tray): support HTTPS server configurations by @abn in #2815
• ci: improve persistent Windows runner hygiene and process lock recovery by @abn in #2837
• fix(build): fix system cpp-httplib detection and HTTPS compilation by @abn in #2814
• fix(custom_args): preserve repeatable flags by @sujikathir in #2840
• fix(installer): remember custom install location by @sujikathir in #2855
• test(ci): wait for server after host rebind by @fl0rianr in #2870
• ci: make C++ test selection merge-friendly by @fl0rianr in #2829
• [bench] Refactor & add hardware info by @bitgamma in #2866
• fix(llamacpp): validate device-backend compatibility during model load by @abn in #2839
• document multi checkpoint model creation by @ckuethe in #2844
• FIX(ci-test-router): recover semantic routing after transient embed failure by @fl0rianr in #2871
• Moving from Quick Rules/ Advanced Rules to Natural Language Router now gives a warning that the progress will be cleared by @meghsat in #2867
• fix(docs): no support for llama.cpp:cpu, metal is used by @fl0rianr in #2881
• Update llama.cpp to b10241 by @github-actions[bot] in #2884
• Bump project version from 11.5.1 to 11.5.2 by @kenvandine in #2894

Full Changelog: v11.5.1...v11.5.2

---

Windows installers are signed. Free code signing provided by SignPath.io, certificate by SignPath Foundation. See our Code Signing Policy.

v11.5.1

Headline

• lemonade bench adds an image-generation benchmark mode with capability-aware scenario and model filtering, response capture, and a new --timeout flag.
• The Router Builder gains a Test Prompt tab, backed by a new POST /routing/validate endpoint, that runs a routing policy against a sample prompt and shows the decision, a step-by-step trace, and a decision-tree view.
• Ten MiniCPM text and vision GGUF models from ModelScope join the built-in llama.cpp catalog.
• Tool-calling requests to llama.cpp with large JSON schema bounds are now accepted, working around a grammar limit that previously rejected valid tool calls.

Breaking Changes

• llama.cpp non-streaming responses now return the requested or registered model id in the model field instead of the local .gguf absolute path.
• Registered and imported collections (user.*/extra.*) now list under their canonical prefixed id on /v1/models, Ollama /api/tags, and MCP lemonade_list_models.
• auto backend selection now prefers Vulkan on AMD systems across all engines, matching the llamacpp behavior.

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_11.5.1-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_11.5.1-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-11.5.1-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-11.5.1-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-11.5.1-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-11.5.1-fc44.aarch64.rpm
macOS	Lemonade-11.5.1-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-11.5.1-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-11.5.1-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-11.5.1-windows-x64.zip
macOS arm64	lemonade-embeddable-11.5.1-macos-arm64.tar.gz

---

What's Changed

Thanks @QuMuzafferEge, @QuTosh, @SlawomirNowaczyk, @ZMXJJ, @abn, @ckuethe, @coolduderules, @fl0rianr, @jeremyfowers, @kenvandine, @ramkrishna2910, @sujikathir, @superm1 for your awesome contributions to this release!

Click to expand changelog

• Update trellis checksums to the pinned v0.4.3 assets by @ramkrishna2910 in #2778
• web: community links in navbar hamburger + homepage hero by @jeremyfowers in #2784
• Bench imagegen by @ckuethe in #2683
• Add libgomp runtime dependency to deb/rpm packages by @kenvandine in #2794
• Fix: Repeated model update reports after pull by @fl0rianr in #2715
• [Router] Fix DirectoryWatcher missing policy delete+rewrite on Windows by @SlawomirNowaczyk in #2798
• Fix llama.cpp model field echoing local .gguf path by @SlawomirNowaczyk in #2796
• Add JSON schema checking to pre-commit by @superm1 in #2802
• fix(cli): parse configuration values as JSON arrays and objects by @abn in #2816
• Fix JSON pre-commit validation with permissive schema by @superm1 with @Copilot in #2810
• fix(http): add libcurl protocol fallback for macOS libcurl compatibility by @abn in #2806
• fix(custom_args): treat negative numbers as values, not flags by @coolduderules in #2773
• Issue 2651: Router Prompt Debugger (Ege Qutwo) by @QuMuzafferEge in #2726
• Add conformance tests (phase 1: deterministic) by @QuTosh in #2761
• Fix: list router/collection ids with their user./extra. prefix in GET /v1/models by @SlawomirNowaczyk in #2797
• Add ModelScope registry search blog post by @ramkrishna2910 in #2828
• Add MiniCPM GGUF models to catalog by @ZMXJJ in #2825
• fix(server): preserve WebSocket auth state and support desktop client origins by @abn in #2790
• Fix Whisper realtime test isolation and stale session work by @fl0rianr in #2821
• docs(telemetry): improve telemetry and arize phoenix setup guide by @abn in #2817
• fix(llamacpp): sanitize oversized tool schema limits by @sujikathir in #2807
• fix(ci): job-isolated process cleanups and fast installer teardowns by @abn in #2643
• Apply a canonical backend preference order to all descriptors by @jeremyfowers in #2835
• wg: Add an Enterprise Grade working group and archive Omni Models by @jeremyfowers in #2812
• Bump project version from 11.5.0 to 11.5.1 by @jeremyfowers in #2849
• Update flm NPU version to v0.9.46 by @jeremyfowers in #2847
• fix(installer): handle slow tarball listings by @sujikathir in #2830

New Contributors

• @QuMuzafferEge made their first contribution in #2726
• @QuTosh made their first contribution in #2761
• @ZMXJJ made their first contribution in #2825
• @sujikathir made their first contribution in #2807

Full Changelog: v11.5.0...v11.5.1

---

Windows installers are signed. Free code signing provided by SignPath.io, certificate by SignPath Foundation. See our Code Signing Policy.

v11.5.0

Headline

• The Lemonade Router is complete: name a collection.router model and requests are steered by rule, classifier, semantic-similarity, or LLM-as-router policies, with per-candidate decision traces, a GUI for authoring and editing routers, and a POST /v1/classify text-classification endpoint backed by a new onnxruntime backend.
• ModelScope is now searchable alongside Hugging Face in the Model Manager, with source-aware model discovery, download, and registration.
• A server-side job engine lets clients post multi-step recipes and manage them through /jobs endpoints, with pause, interrupt, resume, and delete lifecycle control.
• lemond can act as an MCP client host, connecting to external stdio MCP servers and discovering their tools through admin-gated /internal/mcp/* endpoints.
• The lemonade CLI now supports HTTPS/TLS connections to secure remote endpoints, honoring http(s):// schemes in LEMONADE_HOST.

Breaking Changes

• The server no longer sends Access-Control-Allow-Origin: * by default: non-loopback browser origins are now rejected with 403 unless listed in LEMONADE_ALLOWED_ORIGINS, and the WebSocket same-origin fallback has been removed. Remote or LAN browser clients must set LEMONADE_ALLOWED_ORIGINS.

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_11.5.0-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_11.5.0-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-11.5.0-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-11.5.0-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-11.5.0-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-11.5.0-fc44.aarch64.rpm
macOS	Lemonade-11.5.0-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-11.5.0-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-11.5.0-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-11.5.0-windows-x64.zip
macOS arm64	lemonade-embeddable-11.5.0-macos-arm64.tar.gz

---

What's Changed

Thanks @OussamaBATOUCHE, @QuManu, @SlawomirNowaczyk, @abn, @ckopanos, @ekkoitac, @fl0rianr, @jon-hotaisle, @kenvandine, @kyuubyN, @meghsat, @midu16, @pwilkin, @radhaksri, @ramkrishna2910, @superm1, @therealjuanmartinez for your awesome contributions to this release!

Click to expand changelog

• Update version to prepare for release by @superm1 in #2702
• fix(security): pin SHA-256 checksums for contributor-hosted backend assets by @superm1 in #2710
• fix(ci): always run macOS C++ unit tests regardless of signing by @superm1 in #2712
• docs (fix): typo in readme by @jon-hotaisle in #2701
• feat(router): onnxruntime classification backend + /v1/classify (capability for #2592; policy wiring in #2384) by @ramkrishna2910 in #2626
• Fix: re-pulling a collection.router model fails with "unknown key 'source'" by @SlawomirNowaczyk in #2703
• feat: add Modelscope search by @fl0rianr in #2716
• fix(websocket): allow Windows Tauri WebView2 origin for upgrades by @superm1 in #2721
• fix: quote /MANIFESTINPUT path and detect RX 7600 as gfx110X by @kyuubyN in #2678
• Add server-side MCP client host foundation by @fl0rianr in #2605
• fix(docker): Add libreadline8 to Dockerfile dependencies by @ckopanos in #2733
• fix(security): apply HTTP trust policies to POST request paths by @fl0rianr in #2741
• feat: job engine by @pwilkin in #2718
• Keep sd-cpp validation images out of auto-update PRs by @fl0rianr in #2502
• feat(server): support PNA CORS, custom allowed origins, and custom headers by @abn in #2660
• feat(cli): add native HTTPS/TLS and URL scheme parsing support by @abn in #2373
• feat(trellis): per-request UV atlas method, bump trellis to v0.4.3 by @pwilkin in #2667
• fix(docker): include build-essential in runtime stage for vLLM Triton… by @midu16 in #2749
• Wire collection.router classifiers to Router::classify + validate model capability at parse time by @QuManu in #2727
• [Router] llm router classifier + routing.router desugaring (#2405) by @OussamaBATOUCHE in #2698
• Support gfx950, using the same path as gfx942 by @SlawomirNowaczyk in #2687
• feat(telemetry): honor incoming W3C traceparent (opt-in) by @radhaksri in #2734
• [Router] [UI] Author collection.router in the custom-collection panel (#2406) by @meghsat in #2693
• Router end-to-end demo, integration test, and docs (#2388) by @ramkrishna2910 in #2732
• fix(router): isolate routing helpers from standard model LRU slots by @fl0rianr in #2729
• Update version to 11.5.0 by @ramkrishna2910 in #2772
• fix: detect Arch ggml HIP plugin paths by @ekkoitac in #2760
• api: expose is_busy and is_streaming flags in /health response by @therealjuanmartinez in #2720
• Fix arm64 PPA packages missing bundled web UI by @kenvandine in #2770
• fix: resolve CORS blockage on MCP Gateway Endpoint by @abn in #2771

New Contributors

• @jon-hotaisle made their first contribution in #2701
• @ckopanos made their first contribution in #2733
• @midu16 made their first contribution in #2749
• @QuManu made their first contribution in #2727
• @OussamaBATOUCHE made their first contribution in #2698
• @radhaksri made their first contribution in #2734
• @ekkoitac made their first contribution in #2760
• @therealjuanmartinez made their first contribution in #2720

Full Changelog: v11.0.0...v11.5.0

---

Windows installers are signed. Free code signing provided by SignPath.io, certificate by SignPath Foundation. See our Code Signing Policy.

Router Milestone Contributors

The Lemonade Router shipped across v11.0.0 and v11.5.0 (17 PRs). PR authors: @eddierichter-amd, @SlawomirNowaczyk, @ramkrishna2910, @OussamaBATOUCHE, @QuManu, @meghsat, @fl0rianr, @coolduderules. Reviewers: @fl0rianr, @SlawomirNowaczyk, @eddierichter-amd, @ramkrishna2910, @meghsat, @superm1, @bong-water-water-bong.

v11.0.0

Headline

• A new 3D-generation modality ports the Trellis.2 image-to-3D pipeline, exposing a POST /v1/3d/generations endpoint and a GUI 3D panel with an in-app model viewer.
• Text-to-speech arrives with the OpenMOSS backend, including voice-cloning and voice-design models and a dedicated TTS panel in the GUI.
• Naming a collection.router model on /chat/completions, /completions, or /responses now runs its routing engine and reports the selected route via an x_lemonade_route body field and x-lemonade-route header.
• ModelScope is now supported as a second remote model registry alongside Hugging Face, with source selection across the CLI, API, and desktop app.
• The FastFlowLM NPU backend now auto-installs on Linux instead of requiring a manual system-package install, with a new prefer_system config flag.

Breaking Changes

• The Docker runtime image now runs lemond as unprivileged UID 10001 and stores caches under /opt/lemonade/.cache instead of /root/.cache; chown or recreate existing root-owned named volumes.
• Re-registering a Hugging Face user model under a name already used by a different checkpoint now returns an error instead of silently overwriting the existing user.* model; choose a different name.

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_11.0.0-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_11.0.0-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-11.0.0-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-11.0.0-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-11.0.0-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-11.0.0-fc44.aarch64.rpm
macOS	Lemonade-11.0.0-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-11.0.0-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-11.0.0-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-11.0.0-windows-x64.zip
macOS arm64	lemonade-embeddable-11.0.0-macos-arm64.tar.gz

---

What's Changed

Thanks @SlawomirNowaczyk, @abn, @ckuethe, @eddierichter-amd, @fl0rianr, @ianbmacdonald, @indistinctTalk, @kenvandine, @kyuubyN, @pwilkin, @ramkrishna2910, @superm1, @wariobot09 for your awesome contributions to this release!

Click to expand changelog

• Update FLM to 0.9.44 and enable portable FLM for Linux by @superm1 in #2544
• dev: change base image to Ubuntu 24.04 in Dockerfile by @fl0rianr in #2578
• fix(server): replace dummy bind with connect check for port availability by @abn in #2576
• Router/2385 collection.router recipe dispatch by @SlawomirNowaczyk in #2558
• feat(telemetry): add WebSocket server endpoint and security controls by @abn in #2574
• feat(rocm): enable AMD Instinct MI300X (gfx942) in the llama.cpp ROCm backend by @ianbmacdonald in #2587
• Asset generation: OpenMOSS by @pwilkin in #2580
• Start an abstraction layer for benchmark types by @ckuethe in #2557
• Asset generation: Trellis by @pwilkin in #2581
• fix(backends): enable sd-cpp CUDA support on Windows by @kenvandine in #2601
• refactor(audio-gen): apply backend review feedback to ThinkSound and ACE-Step by @pwilkin in #2594
• fix(rocm): detect version-suffixed HIP runtime on Windows (ROCm 7.x) by @ramkrishna2910 in #2603
• Router/2386 decision trace response by @eddierichter-amd in #2606
• fix(docker): drop to non-root user and remove unverified FLM download by @superm1 in #2617
• fix(llamacpp): detect HIP plugin next to PATH-resolved llama-server by @pwilkin in #2628
• fix: model update startup behavior in offline mode by @fl0rianr in #2612
• fix(server): size HTTP thread pool from CPU count and add socket time… by @superm1 in #2618
• log: reduce verbosity of model already loaded and websocket connection logs by @abn in #2599
• Bench embed by @ckuethe in #2577
• fix(auto-label): refactor community user checks to exclude bots by @fl0rianr in #2479
• security: pin FetchContent deps to commit SHAs by @superm1 in #2633
• fix: avoid HF user model name collisions act only when they occur by @fl0rianr in #2058
• security(renderer): redact API key values from console logs by @superm1 in #2632
• fix(acestep): music without lyrics isn't the greatest by @pwilkin in #2629
• fix: handle AMD APU memory pressure correctly by @fl0rianr in #2638
• chore(ci): parallelize link checking, headless setup flags, and documentation link fixes by @abn in #2646
• chore(ci): enable package caching and ccache on distro builds by @abn in #2645
• fix: readme and gen_backend_boilerplate by @fl0rianr in #2658
• ci: add safe AMD64 and ARM64 container validation by @fl0rianr in #2637
• Feat/dxgi use native DXGI QueryVideoMemoryInfo for Windows VRAM pressure by @kyuubyN in #2648
• chore: bump FLM to v0.9.45 by @wariobot09 in #2661
• fix: bump openmoss to v0.1.3 (fixes progressive audio degradation past 10 s) by @pwilkin in #2666
• fix(server): forward ctx_size to auto-loaded models by @indistinctTalk in #2664
• feat(vllm): stage AMD Instinct MI300X (gfx942) support (pending release asset) by @ianbmacdonald in #2588
• feat: make model update checks manual configerable by @fl0rianr in #2659
• docker: add xz support and validate backend extraction by @fl0rianr in #2671
• fix(security): harden HTTP redirects by trust boundary by @superm1 in #2634
• feat(flm): restore validated backend arguments by @fl0rianr in #2677
• fix(security): replace system() with ProcessManager in archive extrac… by @superm1 in #2627
• chore: normalize the tree with pre-commit run --all-files by @ianbmacdonald in #2690
• fix(security): move WebSocket API key from query string to Sec-WebSoc… by @superm1 in #2636
• fix(process): detect zombie backend processes by @fl0rianr in #2697
• feat: add modelscope besides huggingface by @fl0rianr in #2647
• Rename process_unix files to process_linux by @superm1 in #2692

New Contributors

• @kyuubyN made their first contribution in https://github.com/lemonade-sdk/lemonade/pull/...

v10.10.0

Headline

• Audio generation arrives with the ThinkSound sound-effects and ACE-Step music backends, a new POST /v1/audio/generations endpoint, and a GUI audio-generation panel.
• System-installed ROCm is now reused on Windows as well as Linux, and the experimental vLLM ROCm backend reliably targets AMD GPUs on hybrid AMD/NVIDIA hosts.
• A suspend inhibitor keeps the machine awake while inference is running on Linux (systemd-logind), enabled by default and opt-out.
• A new GET /v1/models/{id}/files endpoint reports each model's resolved local files for the GUI Files tab, with absolute paths gated behind an opt-in include_paths=true.

Breaking Changes

• The experimental ROCm vLLM backend now requires a writable $XDG_RUNTIME_DIR; launch fails if one cannot be created, which may affect minimal or container setups.
• After the Debian/Ubuntu package upgrade migration runs, re-enable the per-user service manually with systemctl --user enable lemond.

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_10.10.0-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_10.10.0-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-10.10.0-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-10.10.0-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-10.10.0-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-10.10.0-fc44.aarch64.rpm
macOS	Lemonade-10.10.0-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-10.10.0-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-10.10.0-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-10.10.0-windows-x64.zip
macOS arm64	lemonade-embeddable-10.10.0-macos-arm64.tar.gz

---

What's Changed

Thanks @SlawomirNowaczyk, @abn, @blackdeathdrow, @coolduderules, @eddierichter-amd, @fl0rianr, @jeremyfowers, @kenvandine, @pwilkin, @sofiageo, @superm1, @volen-silo, @wariobot09 for your awesome contributions to this release!

Click to expand changelog

• feat(rocm): detect externally-installed ROCm on Windows by @volen-silo in #2508
• Router/2382 RoutingPolicyEngine — assembly + evaluate() (M8) by @SlawomirNowaczyk in #2499
• docs: add Telemetry guide to mkdocs nav by @jeremyfowers in #2514
• fix(vLLM): ensures vLLM sees only ROCm devices by @fl0rianr in #2007
• docs: update arch linux instructions by @sofiageo in #2517
• fix(installer): add missing architecture_defaults.json to MSI by @blackdeathdrow in #2525
• fix(app): store Windows favicon as regular file by @fl0rianr in #2526
• Add support for inhibiting suspend while inference is running by @superm1 in #2515
• fix: llama.cpp backend updates make sure artifacts are present by @fl0rianr in #2518
• feat: add model files endpoint by @fl0rianr in #2437
• docs: cover tests and API docs in adding-a-backend guide by @jeremyfowers in #2541
• Fix: flaky Whisper realtime transcription event handling by @fl0rianr in #2535
• Fix disabling systemd user service by @superm1 in #2516
• fix: support nested schemas and handle client disconnects in responses telemetry by @abn in #2532
• feat(telemetry): reduce verbosity of telemetry logging by @abn in #2545
• fix(tray): use /live instead of /api/v1/health for server readiness check by @blackdeathdrow in #2523
• docs: add two new YouTube videos to news feed by @jeremyfowers in #2534
• test: relax streaming response length assertions from >5 to >0 by @coolduderules in #2556
• Router/2383 policy parser hot reload by @eddierichter-amd in #2519
• fix(router): capture load duration before std::move to prevent null deref by @coolduderules in #2555
• Enhance contribution guidelines for PR reviews and AI use by @fl0rianr in #2550
• fix: scan entire install_dir for executable permissions after extraction by @wariobot09 in #2549
• feat(backends): audio generation — ThinkSound (SFX) + ACE-Step (music) by @pwilkin in #2530
• feat(router): wire classifier services to router calls by @eddierichter-amd in #2548
• chore: Version bump v10.10.0 in preparation for release by @kenvandine in #2582

New Contributors

• @coolduderules made their first contribution in #2556
• @wariobot09 made their first contribution in #2549

Full Changelog: v10.9.0...v10.10.0

---

Windows installers are signed. Free code signing provided by SignPath.io, certificate by SignPath Foundation. See our Code Signing Policy.

v10.9.0

Headline

• ARM64 (aarch64) Linux packages now ship for Debian, Fedora, the portable embeddable archive, and Docker.
• Users can set custom system prompts for Lemonade omni models, as demonstrated in the new RPG-HaloTales-V1 narrative experience model.
• A new OTLP telemetry subsystem exports traces using OpenInference and OpenTelemetry GenAI conventions, driven by a lemonade telemetry command.
• The vLLM backend now works with Claude Code through the Anthropic Messages API, and adds Qwen3.6 and GLM-4.7-Flash models.

Breaking Changes

• The Linux system tray is now disabled by default (REQUIRE_LINUX_TRAY=OFF), so Linux server packages ship headless; rebuild with the tray option enabled to restore it.

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13 (x86_64)	lemonade-server_10.9.0-debian13_amd64.deb
Debian 13 (ARM64)	lemonade-server_10.9.0-debian13_arm64.deb
Fedora 43 (x86_64)	lemonade-server-10.9.0-fc43.x86_64.rpm
Fedora 43 (ARM64)	lemonade-server-10.9.0-fc43.aarch64.rpm
Fedora 44 (x86_64)	lemonade-server-10.9.0-fc44.x86_64.rpm
Fedora 44 (ARM64)	lemonade-server-10.9.0-fc44.aarch64.rpm
macOS	Lemonade-10.9.0-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-10.9.0-ubuntu-x64.tar.gz
Ubuntu arm64	lemonade-embeddable-10.9.0-ubuntu-arm64.tar.gz
Windows x64	lemonade-embeddable-10.9.0-windows-x64.zip
macOS arm64	lemonade-embeddable-10.9.0-macos-arm64.tar.gz

---

What's Changed

Thanks @GabrielReusRodriguez, @SlawomirNowaczyk, @abn, @bitgamma, @blackdeathdrow, @bong-water-water-bong, @ckuethe, @clemperorpenguin, @eddierichter-amd, @fl0rianr, @gchriswill, @github-actions, @ianbmacdonald, @jeremyfowers, @kenvandine, @lucifer-vali, @ramkrishna2910, @sawansri, @siavashhub, @superm1, @volen-silo for your awesome contributions to this release!

Click to expand changelog

• fix(packaging): don't auto-enable the per-user lemond.service (Closes #2125) by @ianbmacdonald in #2354
• fix(config): accept flat backend keys like vllm_args in config set (#1824) by @ramkrishna2910 in #2394
• test/ci: harden CLI endpoint suites against shared-server flakes by @fl0rianr in #2365
• Lemonade cli list models with size in Gb by @GabrielReusRodriguez in #2329
• Handle separate sd-cpp CUDA release pins by @fl0rianr in #2369
• fix: refuse to start a second lemond when the port is already in use (#2255) by @siavashhub in #2258
• Fix sd-cpp:metal install (resolve asset filename wildcard) by @jeremyfowers in #2402
• Allow HTTP custom backends with warnings by @sawansri in #2370
• Add comment and PR description guidelines to AGENTS.md by @superm1 in #2409
• [Router] Interfaces + fixtures stub (foundation) by @ramkrishna2910 in #2408
• fix: Prevent the Rosetta prompt on macOS installation by @clemperorpenguin in #2297
• fix(cli): resolve option inheritance bug in launch help options by @abn in #2396
• Move omni collection system prompt to per-model JSON by @ramkrishna2910 in #2395
• fix: ROCm Support for RDNA2 Device Models by @gchriswill in #2410
• feat: expand system stats with storage infos by @fl0rianr in #2436
• test: fix whisper realtime macOS VAD flake by @abn in #2442
• test: fix stale docstring and add VAD rationale comment by @bong-water-water-bong in #2447
• feat(telemetry): add unified OTLP telemetry with OpenInference and OpenTelemetry GenAI support by @abn in #2330
• Router/2378 2379 evaluator registry by @eddierichter-amd in #2440
• test/ci: follow up on hosted Ollama tool-calling timeouts by @fl0rianr in #2399
• website: Overhaul the front page by @jeremyfowers in #2427
• refactor(backends): self-describing WrappedServer backends (#2287) by @jeremyfowers in #2320
• Fix startup self-deadlock with a thread-local recursion guard (minimal alt to #2443) by @jeremyfowers in #2488
• fix(launch): resolve npm .cmd launcher on Windows for pi and codex by @jeremyfowers in #2489
• Fix three broken doc links found by zensical by @lucifer-vali in #2486
• Enable vLLM + Messages API and per model family vLLM args by @sawansri in #2143
• [Router] 2381 unified interface for semantic_similarity and model classifier by @SlawomirNowaczyk in #2483
• Fix flaky Ollama tool-calling CI by disabling default thinking for tool requests by @fl0rianr in #2485
• Merge release-v10.8.1 back to main for proper release tag handling by @kenvandine in #2493
• fix(backends): fallback to installed backend when offline by @fl0rianr in #2393
• Fix (#2412) : Added offline check before pulling a model on server.cpp by @GabrielReusRodriguez in #2434
• Add arm64 builds to all Linux release packages and PPA by @kenvandine in #2430
• fix(rocm): resolve backend assets to published gfx family names (#2415) by @superm1 in #2490
• Add lemonade and lemond options to bug report template by @ckuethe in #2444
• add per-architecture options by @bitgamma in #2496
• Publish rolling :edge and immutable :sha- container images on merge to main by @siavashhub in #2397
• Detect externally-installed ROCm before downloading TheRock by @volen-silo in #2498
• docs(release): add step to reconcile release tag into main by @jeremyfowers in #2504
• fix(server): recover stalled HF model downloads by @fl0rianr in #2421
• [Router] Deterministic conditions: keywords / regex / min_chars / has_* / metadata (M4, #2380) by @ramkrishna2910 in #2491
• fix(packaging): strip lemonade-tray from RPM by @lucifer-vali in #2484
• Update stable-diffusion.cpp to master-721-8caa3f9 and CUDA master-726-db48014 by @github-actions[bot] in #2477
• feat: detect HuggingFace model updates and show Update button in web UI by @blackdeathdrow in #2243
• docs: list ARM64 Linux release artifacts left out by #2430 by @jeremyfowers in #2505
• chore(release): bump version to v10.9.0 by @jeremyfowers in #2506
• feat(models): add RPG-HaloTales-V1 omni collection by @jeremyfowers in https://github.com/lemonade-sdk/le...

v10.8.1

Headline

• Speculative decoding now accepts draft/MTP/EAGLE3 checkpoints, with new Gemma-4 MTP models ready to pull.
• ROCm installation and GPU detection are restored for Radeon RX RDNA2/3/4 dGPUs on Windows and Linux.
• Backend installs are now crash-safe, with resilient downloads that fall back gracefully when a release lookup or cache snapshot is unavailable.
• lemonade bench --response-log captures each model's responses and run metadata to a JSONL file for later quality evaluation.
• The lemonade backends command now lists only supported recipes and backends by default; use lemonade backends --all to see every available option.

Breaking Changes

• The --model-draft, -md, and --spec-draft-model flags are now reserved for internal speculation-decoding support and can no longer be passed manually through llamacpp_args.

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13	lemonade-server_10.8.1-debian13_amd64.deb
Fedora 43	lemonade-server-10.8.1-fc43.x86_64.rpm
Fedora 44	lemonade-server-10.8.1-fc44.x86_64.rpm
macOS	Lemonade-10.8.1-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-10.8.1-ubuntu-x64.tar.gz
Windows x64	lemonade-embeddable-10.8.1-windows-x64.zip
macOS arm64	lemonade-embeddable-10.8.1-macos-arm64.tar.gz

---

What's Changed

Thanks @GabrielReusRodriguez, @Kushal1213, @Phqen1x, @abn, @bitgamma, @blackdeathdrow, @ckuethe, @fl0rianr, @github-actions, @ianbmacdonald, @jeremyfowers, @jtlayton, @kenvandine, @lucifer-vali, @matthewjhunter, @ramkrishna2910, @sagebind, @superm1 for your awesome contributions to this release!

Click to expand changelog

• ci: support tagging releases from a release branch by @jeremyfowers in #2272
• ci: add repo-manager workflow by @jeremyfowers in #2276
• fix(whisper): drop gfx103X from rocm whisper supported archs by @ramkrishna2910 in #2274
• test(gguf): add unit tests for MTP / capability label detection (#2176) by @ramkrishna2910 in #2281
• fix(rocm): accept wildcard GPU arch families in TheRock install gate (#2093 follow-up) by @ramkrishna2910 in #2280
• fix(test): require Whisper model load before language transcription test by @fl0rianr in #2278
• docs(cli): correct launch note - LEMONADE_* recipe env vars no longer honored by @ramkrishna2910 in #2275
• ci(repo-manager): automatic repo-manager fully operational by @jeremyfowers in #2285
• docs(release): update release process for repo-manager automation by @jeremyfowers in #2288
• fix(ci): restart MacOS server before whisper metal tests by @fl0rianr in #2303
• auto update and validate sd-cpp by @fl0rianr in #2139
• fix(macOS): updates whisper from v1.8.4 to v.1.8.5 for metal by @fl0rianr in #2309
• Fix Problem with .devcontainer. It did not copy .devcontainer/reinstall-cmake.sh when building container by @GabrielReusRodriguez in #2273
• Add support for additional draft checkpoint by @bitgamma in #2317
• docs: fix Debian 13 installation docs for issue #2299 by @superm1 in #2323
• docs: add documentation style guide for community and AI-assisted contributions by @kenvandine in #2054
• Fix copy-to-clipboard buttons silently failing in web-app over HTTP by @blackdeathdrow in #2260
• Fix(cli): Lemonade backends now shows only supported backends . Added --all option to show all backends. by @GabrielReusRodriguez in #2254
• fix(gpu): support wildcards in GPU detection logic by @jtlayton in #2295
• Fix ROCm whisper-server startup: add TheRock lib dir to LD_LIBRARY_PATH by @matthewjhunter in #2293
• Capture output from lemonade bench by @ckuethe in #2214
• fix(windows): use ProcessManager::run_command for 7z extraction instead of system() Fixes #2313 by @Phqen1x in #2322
• fix(server): fall back to installed llama.cpp binary when "latest" release lookup fails by @ianbmacdonald in #2279
• fix(moonshine): return 400 on invalid audio by @abn in #2326
• fix(backends): stage and verify backend install before removing the working binary by @ianbmacdonald in #2315
• docs: system-stats API by @jeremyfowers in #2284
• fix: return 400 instead of 500 when request body is empty or not valid JSON by @Kushal1213 in #2232
• devcontainers - Create the python env and install reqs to allow python test execution by @GabrielReusRodriguez in #2336
• systemd: add user-service symlink at /usr/lib/systemd/user/ by @lucifer-vali in #2173
• fix: resolve shared-repo GGUF variants orphaned by refs/main advance by @ianbmacdonald in #2311
• Fix: eviction_engine does not run nvidia-smi on AMD configs anymore and some testing problems fixed. by @GabrielReusRodriguez in #2331
• Update stable-diffusion.cpp to master-709-92a3b73 by @github-actions[bot] in #2335
• Support image[] parameter for /v1/images/edits by @sagebind in #2321
• Update llama.cpp to b9747 by @github-actions[bot] in #2333
• Better auto context size estimate by @bitgamma in #2337
• Version bump to 10.8.1 to prepare for the release by @kenvandine in #2374

New Contributors

• @jtlayton made their first contribution in #2295
• @matthewjhunter made their first contribution in #2293
• @sagebind made their first contribution in #2321

Full Changelog: v10.8.0...v10.8.1

---

Windows installers are signed. Free code signing provided by SignPath.io, certificate by SignPath Foundation. See our Code Signing Policy.

v10.8.0

Headline

• Live model management: auto-unload is available, along with the ability to pin models to prevent eviction.
• LMX-Omni image generation gains fine-grained parameters, unified collection import/export, and Hugging Face distribution.
• Cloud offload serves chat completions from any OpenAI-compatible provider alongside local models.
• New MCP gateway and lemonade launch pi integrations let external tools and agents call local models.
• Expanded platform support adds Moonshine speech-to-text, NVIDIA GB10 arm64, Debian 13, and ROCm for Radeon GPUs.

Breaking Changes

• vLLM model IDs renamed from Qwen3.5-*-vLLM to Qwen3.5-*-FP16-vLLM.
• The --flm-args CLI flag and flm_args API parameter have been removed.
• Recipe-configuration environment variables have been removed.
• The legacy GUI collection export bundle format can no longer be imported.
• The default ctx_size changed from 4096 to -1 (auto-tuned).

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Debian 13	lemonade-server_10.8.0-debian13_amd64.deb
Fedora 43	lemonade-server-10.8.0-fc43.x86_64.rpm
Fedora 44	lemonade-server-10.8.0-fc44.x86_64.rpm
macOS	Lemonade-10.8.0-Darwin.pkg

Other platforms? See our Installation Options for Docker, Snap, Arch, Debian, and more.

Embeddable Lemonade

Portable binaries for bundling into your own installer. Run lemond ./ as a subprocess.

Platform	Download
Ubuntu x64	lemonade-embeddable-10.8.0-ubuntu-x64.tar.gz
Windows x64	lemonade-embeddable-10.8.0-windows-x64.zip
macOS arm64	lemonade-embeddable-10.8.0-macos-arm64.tar.gz

---

What's Changed

Thanks @Geramy, @Kushal1213, @abn, @anditherobot, @bitgamma, @blackdeathdrow, @ckuethe, @clemperorpenguin, @fl0rianr, @github-actions, @iswaryaalex, @jeremyfowers, @kenvandine, @neoblizz, @noamsto, @ramkrishna2910, @sawansri, @siavashhub, @superm1, @vgodsoe for your awesome contributions to this release!

Click to expand changelog

• ci: wire HF_TOKEN into linux distro builds; drop Fedora by @jeremyfowers in #2169
• Fix light theme persistence by @anditherobot in #2121
• cli: group run/load recipe options by @anditherobot in #2144
• docs: document pre-v10 legacy Linux config.json path by @anditherobot in #2145
• fix(vllm): report streaming TTFT and TPS by @fl0rianr in #2009
• Fix compat with vscode by @superm1 in #2146
• trivial: Run pre-commit whitespace hook by @superm1 in #2140
• feat(backends): add Moonshine streaming STT backend (#2115) by @Geramy in #2178
• Rewrite heartbeat + monitor workflows by @vgodsoe in #1960
• ci(windows): migrate Windows builds to Visual Studio 2026 by @jeremyfowers in #2199
• fix: remove load options environment variable support from lemonade CLI by @superm1 in #2185
• feat: auto-detect MTP capability from GGUF metadata by @blackdeathdrow in #2176
• moonshine follow-ups: dedupe realtime docs (#2192), GUI label (#2193), platform list by @Geramy in #2196
• fix(ui): connect log stream on main HTTP port instead of dedicated WebSocket port by @blackdeathdrow in #2191
• Add support for launching pi.dev by @superm1 in #2198
• fix: use Win32 API fallback for create_directories on symlinked Huggi… by @superm1 in #2184
• fix: enable TheRock ROCm installation on Windows for Radeon RX GPUs by @anditherobot in #2093
• Add triage dashboard with GitHub Pages auto-publish by @ramkrishna2910 in #2159
• Fix CUDA backend on NVIDIA GB10 (sm_121) arm64 by @kenvandine in #2203
• Refactor codebase to drop so much #ifdef _WIN32 by @superm1 in #2142
• ci: narrow API key CLI smoke tests by @fl0rianr in #2149
• Fix: recipe/backend status resolution for duplicated backend definition by @fl0rianr in #2206
• docs: add LangChain integration guide for Lemonade Server by @Kushal1213 in #2038
• vllm: rename built-in models with -FP16- suffix by @ramkrishna2910 in #2155
• Enable ROCm on gfx1152 by @ckuethe in #2186
• Show models of several backends in agent launch model picker by @sawansri in #2205
• ci: add engine::moonshine to auto-labeling by @ramkrishna2910 in #2209
• ci: grant pull-requests:read to triage dashboard workflow by @ramkrishna2910 in #2210
• Add instructions about dkms conflict for users that run into that by @vgodsoe in #2208
• feat(cli): add interactive chat in the terminal by @siavashhub in #2207
• Fix AMD GPU architecture detection for CDNA datacenter parts by @neoblizz in #2182
• feat(config): honor LEMONADE_DEFAULTS_PATH for non-FHS installs by @noamsto in #2220
• fix(cli): lemonade launch claude fails to launch on Windows, throws Error 193 by @anditherobot in #2213
• Fix download speed and ETA display at very low speeds by @fl0rianr in #2219
• Cloud offload: provider-agnostic OpenAI-compat backend + UI by @ramkrishna2910 in #1785
• Generate an artifact for Debian 13 by @superm1 in #2240
• Add inotify to extra models directory by @superm1 in #2217
• feat: Add backend watchdog for hung child servers by @fl0rianr in #2028
• Update llama.cpp to b9632 by @github-actions[bot] in #2238
• docs(config): document defaults.json seeding and LEMONADE_DEFAULTS_PATH by @noamsto in #2221
• feat: opt-in dynamic VRAM management (auto-evict idle models) by @clemperorpenguin in #2183
• fix(rocm): repair stable runtime when backend is update_required by @fl0rianr in #2003
• Drop support for flm_args by @superm1 in #2015
• fix: re-download missing checkpoint files automatically (no restart needed) by @blackdeathdrow in #2239
• trivial: Update debian/copyright by @superm1 in #2244
• cli: use long read timeout for pull and install streams by @abn in #2222
• docs: add release process guide by @jeremyfowers in #2200
• Fix(Omni): image tool options for Omni image generation by @fl0rianr in #1998
• fix(windows): set safe cwd for backend child processes by @fl0rianr in #2246
• ci: keep Windows inference cache outside workspace by @fl0rianr in #2148
• ci: drop src/cpp from distro build triggers by @jeremyfowers in #2170
• fix: align moonshine backend lifecycle with watchdog by @fl0rianr in #2252
• test: pre-warm ROCm (TheRock) runtime before rocm-backend tests by @jeremyfowers in #2250
• fix(server): han...