v10.7.0

Headline

• LMX-Omni models are now supported for any OpenAI API compatible app that can render multimedia output, including Open WebUI and AnythingLLM.
• The new lemonade bench command provides apples-to-apples LLM benchmarking across llama.cpp, FastFlowLM, vLLM, and Ryzen AI SW.
• Greatly improved cross-platform compatibility: llama.cpp CUDA support added for Windows and Linux; stable-diffusion.cpp CUDA added for Linux, stable-diffusion.cpp Vulkan added for Windows and Linux.
• Added a native Prometheus endpoint for real time stats monitoring.

Breaking Changes

• The deprecated environment variables for lemond configuration have finally been removed, use lemonade config instead.
• Windows AMD users are highly encouraged to update their Adrenalin driver before generating images.

Lemonade Server

Operating System	Downloads
Windows	lemonade.msi
Ubuntu 24.04+	Launchpad PPA
Fedora 43	lemonade-server-10.7.0-fc43.x86_64.rpm
Fedora 44	lemonade-server-10.7.0-fc44.x86_64.rpm
macOS	Lemonade-10.7.0-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-10.7.0-ubuntu-x64.tar.gz
Windows x64	lemonade-embeddable-10.7.0-windows-x64.zip
macOS arm64	lemonade-embeddable-10.7.0-macos-arm64.tar.gz

---

What's Changed

Thanks @Geramy, @clemperorpenguin, @Kushal1213, @Phqen1x, @Storce, @ZaneNi, @anditherobot, @bitgamma, @ckuethe, @fl0rianr, @github-actions, @jeremyfowers, @julianxhokaxhiu, @kenvandine, @kylemanna, @lucifer-vali, @noamsto, @ramkrishna2910, @sawansri, @siavashhub, @superm1 for your awesome contributions to this release!

Click to expand changelog

• Stop backend archive staging from using shared /tmp during ROCm installs by @Copilot in #1996
• Add bench command by @bitgamma in #1953
• Remove more cases of hardcoded /tmp or predictable paths to mitigate … by @superm1 in #1997
• speech: Avoid duplicate Content-Type headers by @kylemanna in #2005
• fix: harden local model path finder by @fl0rianr in #1989
• Add llamacpp:cuda backend for NVIDIA GPUs by @Phqen1x in #1772
• feat: update FLM version to v0.9.43 by @ZaneNi in #2026
• fix(kokoro): correct ESPEAK_DATA_PATH construction by @anditherobot in #2025
• fix: harden markdown link handling against prompt injection by @superm1 in #2016
• security: fix path traversal in NPU cache download by @superm1 in #2017
• fix(readme): row arrangement fixed by @fl0rianr in #2023
• Smooth displayed download speed by @fl0rianr in #1995
• fix: bump macOS sd.cpp metal version and macOS release number by @Geramy in #2030
• Add LFM2.5-8B-A1B by @sawansri in #2036
• Fix downloaded status for multi-checkpoint models by @anditherobot in #2031
• add instructions to start/enable lemond on Fedora based distros by @lucifer-vali in #2029
• Correct websockets not to listen on 0.0.0.0 by default by @superm1 in #2034
• fix(server): translate model names before checking if they are downloaded (#2014) by @anditherobot in #2048
• Map LEMONADE_WHISPERCPP_VULKAN_{BIN,ARGS} env vars for parity by @noamsto in #2045
• Link system cpp-httplib portably (header-only friendly) by @noamsto in #2047
• feat: Add WSL support via librocdxg by @julianxhokaxhiu in #2049
• Vulkan back-end support for sdcpp by @ckuethe in #2039
• Update URL for CUDA backend for llama.cpp (llamacpp:cuda) by @Phqen1x in #2011
• fix: stop refreshing models on download progress by @fl0rianr in #2051
• Allow LEMONADE_GGML_HIP_PATH to locate the HIP plugin on non-FHS installs by @noamsto in #2044
• Fix: Adds SHA check on every download by @fl0rianr in #1950
• Show 3-line descriptions in Marketplace cards by @Copilot in #2069
• Fix: HF snapshot ref reuse for unchanged model artifacts by @fl0rianr in #2066
• fix: user model RAM filtering by @fl0rianr in #2067
• test: give pull multi endpoint test a unique number by @fl0rianr in #2059
• fix(vLLM): make startup timeout adjustable, increased global_timeout by @fl0rianr in #2004
• Fix spurious model re-downloads: make client pulls cache-first by default by @jeremyfowers in #2065
• fix: user model registry writes and stale cache lookups by @fl0rianr in #2064
• fix(cuda): NVIDIA Blackwell/PRIME detection and CUDA device fixes on Linux by @kenvandine in #2070
• security: validate URL scheme in openExternal handler by @superm1 in #2019
• Fix #1959: Replace old Qwen example model in Lemonade API docs by @anditherobot in #2087
• Add Gemma 4 12B by @sawansri in #2090
• feat: server-side Omni orchestration for /chat/completions (Open WebUI) by @jeremyfowers in #2071
• Separate local and available model list output by @anditherobot in #2086
• Create a set of working groups by @jeremyfowers in #2079
• CI: Limit distro builds to relevant changes by @fl0rianr in #2104
• ci: consolidate GitHub-hosted test jobs to reduce download rate limiting by @jeremyfowers in #2105
• Update llama.cpp to b9518 by @github-actions[bot] in #2101
• feat: support benchmarking multiple models in a single command by @ckuethe in #2075
• feat: add CUDA backend support for stable-diffusion.cpp by @Phqen1x in #2062
• Native Prometheus Metrics Endpoint by @Storce in #1897
• Auto-label issues and PRs by backend and area by @ramkrishna2910 in #2076
• ci / test: fix the 409 CI error by @fl0rianr in #2117
• ci: set HF_HOME before starting hosted Linux server by @fl0rianr in #2116
• test: add coverage for GET /v1/pull/variants (4 paths) by @Kushal1213 in #2096
• Drop all environment variable migration code by @superm1 in #2106
• Add Debian 13 to NPU installation instructions by @superm1 in #2114
• Remove pointless comments from codebase by @superm1 in #2133
• Update llama.cpp to b9550 by @github-actions[bot] in #2132
• fix(security): prevent path traversal in /web-app/ route by @superm1 in #2084
• test: retry transient model pull setup failures by @fl0rianr in #2123
• feat: add Linux ARM64 CPU and Vulkan llamacpp backend support by @kenvandine in #2081
• Lemonade CLI within Docker by @siavashhub in #2126
• update sd-cpp backend by @fl0rianr in #2102
• auto-label: teach prompt about area::ci + test coverage ≠ documentation by @ramkrishna2910 in #2129
• fix(sd image): seed UX and drop generation timeout by @fl0rianr in #1924
• fix: remove references to env vars dropped in #2106 by @jeremyfowers in #2152
• Add roadmap for the Auto-Tune working group by @bitgamma in #2138
• Readme: Add cuda and vulkan for image gen by @fl0rianr in #2160
• ci: require stx-halo runner for server_omni.py tests by @jeremyfowers in #2157
• Allow internal endpoints from remote addresses; warn on unsecured non-loopback bind by @jeremyfowers in #2100
• fix(nvidia/snap): CUDA GPU detection under snap strict confinement by @kenvandine in #2154
• add cuda to llama.cpp auto updater by @fl0rianr in #2151
• fix: corrupted dowload speed/eta during finalizing by @fl0rianr in #2165
• Update llama.cpp to b9585 by @github-actions[bot] in #2168
• Adjust runtime directory handling for systemd awareness by @clemperorpenguin and @superm1 in #2167
• chore: bump version to 10.7.0 by @jeremyfowers in #2172
• Remove sd-cpp cuda Windows support from compatibility tables by @jeremyfowers in #2174

New Contributors

• @clemperorpenguin made their first contribution in #2167
• @kylemanna made their first contribution in #2005
• @noamsto made their first contribution in #2045
• @julianxhokaxhiu made their first contribution in #2049
• @Storce made their first contribution in #1897
• @Kushal1213 made their first contribution in #2096

Full Changelog: v10.6.0...v10.7.0

---

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