Skip to content

Release Announcements

Jump to bottom
Jeremy Fowers edited this page Jun 25, 2026 · 3 revisions

Each release is announced on Discord and copied here for posterity.

Format: put a L1 heading # with the exact content of the git tag for the release, then paste the content of the announcement verbatim below that.

v10.8.1

Lemonade v10.8.1

@releases Fresh Lemonade is out! This release brings you the power of speculative decoding with new Gemma-4 MTP models, gets your Radeon cards back on track, and makes backend installs tougher than ever — plus a neat bench logging trick for measuring output quality.

Breaking Changes

  • The --model-draft, -md, and --spec-draft-model flags are now reserved for internal speculative-decoding use, so they can no longer be slipped through llamacpp_args.

Speculative Decoding with Gemma-4 Models

@bitgamma added support for draft (drafter/MTP/EAGLE3) checkpoints in the llamacpp recipe and registered a batch of new Gemma-4 MTP models that you can pull and try. The automatic ctx_size heuristic was also refactored to weight KV-cache cost across sliding-window vs full-attention layers, giving every platform — not just one OS — more accurate context estimates.

Radeon RX dGPUs are back on track

If you're rocking a RDNA2/3/4 Radeon card, ROCm installs should work again thanks to <@924840864432390165> (wildcard arch support in the install gate) and @jtlayton (GPU detection prefix-wildcard fix for KFD-derived arches like gfx1103). And a shout-out to @matthewjhunter for prepending LD_LIBRARY_PATH when launching the ROCm whisper-server so it no longer aborts at startup over libamd_comgr.so.3.

Crash-safe backend installs

@ianbmacdonald made (re)installations resilient: downloads and extraction happen in a staging directory first, and the binary only swaps into place once verified — so an interrupted download can never leave you with a broken backend anymore. If your GitHub "latest" lookup is unreachable, Lemonade falls back to your installed version instead of blowing up.

Quality metrics with bench --response-log

@ckuethe added --response-log FILE to lemonade bench, so every model's response gets captured alongside run metadata into a JSONL file — perfect for post-run quality evaluation or side-by-side comparisons.

Additional Improvements

  • Better HTTP error handling: bad audio files and malformed request bodies now return proper 400s instead of leaking 500s, by @abn and @Kushal1213.
  • The /v1/images/edits endpoint accepts image[] array syntax (e.g. Open WebUI compatibility), by @sagebind; copy-to-clipboard buttons work over plain HTTP now, by @blackdeathdrow.
  • Windows 7z extraction no longer flashes a console window — clean heads-up display thanks to <@239445513542303755>.
  • CI automation keeps stable-diffusion.cpp backend pins healthy and auto-update-ready, by @fl0rianr; macOS Metal whisper got bumped to v1.8.5 (then dialed back to v1.8.4 for a known issue) with test hardening from @fl0rianr and @GabrielReusRodriguez.
  • Docs polish: Debian 13 install fix by @superm1, plus a brand-new documentation style guide by <@379811340564299777> to keep future contribs on point.
  • lemon-prod pool-membership labels now gate all self-hosted runner jobs in the CI/release workflows, with fresh docs from @jeremyfowers covering the release-branch workflow end-to-end.
  • lemonade backends now shows only recipes and backends we actively support by default; run lemonade backends --all if you want to see the complete list.

Full release notes: https://github.com/lemonade-sdk/lemonade/releases/tag/v10.8.1 — let us know how those Radeon fixes work for you!

v10.8.0

Lemonade v10.8.0

@everyone Big one today — Moonshine speech-to-text, Model Management improvements, Cloud offload, the MCP gateway, and a whole lot of platform love. Let's go.

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 are gone (security hardening).
  • Recipe environment variables removed.
  • Legacy GUI collection export bundle format no longer imports.
  • The default ctx_size is now -1 (auto-tuned) instead of 4096.

🎙️ Moonshine On-Device Speech-to-Text

<@336691752645689345> , <@372211157735309313>, @Theohox, and <@1440101722365362440> all put work into Moonshine, and the result is a streaming speech-to-text backend you can hit over HTTP, TCP JSONL, or the OpenAI-compatible WebSocket realtime endpoint. It's on-device, it's fast, and it ships with Tiny, Small, and Medium model variants.

🧠 Model Management Overhaul

<@189761363529170945> and <@341765612478922755> built dynamic VRAM memory management that automatically unloads idle models and downsizes their KV-cache to reclaim GPU memory on the fly, plus model pinning so the models you want kept hot never get evicted. <@569452289195114502> added automatic context size determination, so Lemonade picks the right context length from your available system memory and the model's architecture, no manual tuning required.

🖼️ LMX-Omni Image Improvements + Unified Collections

<@1440101722365362440> exposed fine-grained image parameters - size, steps, cfg_scale, seed, sample method, and flow shift, so you have real control over Omni image generation, and <@1354802563706519594> with <@1440101722365362440> unified the collection import/export format across server, GUI, CLI, and Hugging Face. Pull and share LMX-Omni collections directly from HF now.

☁️ Cloud Offload

<@924840864432390165> built a provider-agnostic cloud offload backend so you can serve chat completions from any OpenAI-compatible provider Fireworks, OpenRouter, Together, OpenAI right alongside your local models. New CLI subcommands and a rewritten UI panel make it dead simple to register and switch providers.

🖥️ Platform Support

<@379811340564299777> added NVIDIA GB10 (Blackwell) arm64 CUDA support, <@1138231689525600326> with @superm1 enabled TheRock ROCm on Windows for Radeon RX GPUs, @superm1 added a dedicated Debian 13 build, @ckuethe added ROCm for the Radeon 840M/860M (gfx1152), <@749492457510666362> moved whisper.cpp to ROCm on Windows and Linux, and @neoblizz fixed CDNA datacenter GPU architecture detection.

Additional Improvements

  • Interactive terminal chat REPL (lemonade chat) with slash commands, reasoning support, and cross-platform signal handling — by @siavashhub.
  • Backend watchdog that detects and recovers from hung or crashed child servers — by <@1440101722365362440>.
  • Automatic re-download of missing model checkpoint files without requiring a server restart — by @blackdeathdrow.
  • CLI help output grouped by backend for much easier flag discovery — by <@1138231689525600326> .
  • Fixed misleading download speed and ETA display near the end of downloads — by <@1440101722365362440>.
  • Model Manager "Downloaded only" toggle now persists across app restarts — by <@1138231689525600326>.

Full release notes are up on GitHub, let me know what you think!

v10.7.0

Lemonade v10.7.0

@everyone we have a massive release for you today!

News

Lemonade has a bi-weekly public maintainers meeting now. The last one was June 5, with the notes available here and the next one is June 19.

  • We have a real roadmap now, focused around 6 areas: omni models, auto-tuning, cross-vendor support, cloud-hybrid usage, remote usage, and overhauling the GUI app.
  • We are moving to a weekly release cadence, each Wednesday morning, starting today!

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.

Omni Models

LMX-Omni models are now supported in any OpenAI API compatible app that can render multimedia output! Check out the screenshot below. Thanks <@1440101722365362440> for teaming up to make this a reality!

Auto Tuning

You can't optimize what you can't measure. That's why <@569452289195114502> <@416634261961965579> <@589490993091313706> and everyone in the working group have introduced the lemonade bench command, which provides apples-to-apples LLM benchmarking across llama.cpp, FastFlowLM, vLLM, and Ryzen AI SW. Check it out here!

Cross-Vendor Support

<@379811340564299777> <@239445513542303755> <@416634261961965579> <@1440101722365362440> and more are working to give Lemonade excellent support on more platforms. This release brings:

  • llama.cpp CUDA on Windows and Linux, and stable-diffusion.cpp CUDA on Linux
  • stable-diffusion.cpp Vulkan on Windows and Linux
  • build support for non-Apple ARM64 and llama.cpp Vulkan for Adreno GPUs

This means LMX-Omni models are now GPU accelerated on AMD, Nvidia, Intel, and Apple systems 🚀

Additional Improvements

This release had over 70 contributions from 19 authors and 9 reviewers!

  • Native Promethus metrics endpoint by <@478620327681720320> and @sbates130272
  • WSL ROCm support by <@381032614271909890>
  • Greatly improved download experience for models and backends, including SHA checks, by <@1440101722365362440>
  • Security overhaul (powered by Mythos Preview!) by <@372211157735309313> <@569452289195114502> <@1440101722365362440>
  • CI system overhaul by myself and <@1440101722365362440> to make contributing more fun
  • /internal endpoints now accessible by remote connections by me and lemonade CLI available inside of Dockers thanks to <@1047320580346626088>, should make the Docker experience better!
  • So much more, please see the full release notes: https://github.com/lemonade-sdk/lemonade/releases/tag/v10.7.0

HUGE THANK YOU to everyone who contributed and reviewed, you're what makes Lemonade great ❤️

v10.6.0

Lemonade v10.6

@everyone we have new fresh Lemonade for you to enjoy this weekend!

Omni Model Glow Up

<@1440101722365362440> and I have put a fresh coat of paint on OmniRouter, now simply called Lemonade Omni Models.

  • Add your own custom omni models in the GUI (File > New Omni Model) and CLI. Full guide here.
  • Upgraded the omni models with Qwen3.6 and MTP, and adopted a new naming scheme called Lemonade Mix (LMX):
    • LMX-Omni-52B-Halo is perfect for high-RAM devices like Strix Halo with Qwen3.6-35B MTP MoE, Flux Klein image editing.
    • LMX-Omni-5.5B-Lite delivers true omni-modality on a RAM budget with Qwen3.5-4B MTP and SD-Turbo image gen.

Additional Improvements

  • Greatly improved download robustness by <@1440101722365362440> (you can safely close your browser tabs during downloads now!)
    • This may have fixed thorny issue 1967, please let us know!
  • A trio of fixes for Linux by <@341765612478922755> , one for macOS by <@336691752645689345>, and one for error reporting by neubig
  • Custom downloaded models now show the correct size on disk thanks to <@589490993091313706> and correctly apply the vision and tool-calling labels thanks to <@1440101722365362440>

Check out the release notes here and a quick screenshot of the omni model refresh below!

v10.5.1

Lemonade v10.5.1

@everyone a quick release today to cover two important things:

MTP is supported out-of-box

lemonade pull Qwen3.6-27B-MTP-GGUF

Run it with ROCm 7.13: lemonade load Qwen3.6-27B-MTP-GGUF --llamacpp rocm

We set the llamacpp args for you to make things easy, but you're welcome to override with your own settings.

Thanks <@589490993091313706> <@569452289195114502> <@336691752645689345> and others for test driving this before release!

Fedora 43 is back!

Thanks <@1239555227011780709> for jumping in! Now we support separate RPMs for Fedora 43 and 44, both available on the release page.

Additional Improvements

  • Set args on a per-backend (e.g., vulkan vs. rocm) basis by <@569452289195114502>
  • Custom model support for stable-diffusion.cpp models by <@1440101722365362440>
  • We're now on upstream stable-diffusion.cpp + ROCm builds! by <@372211157735309313>

And more, please see the full release notes!

v10.5.0

Lemonade v10.5

Another week, another big release @everyone! This one brings ROCm 7.13, full macOS support, and much more.

🪨 ROCm 7.13

This is the first "stable" release of ROCm based on TheRock and <@372211157735309313> has brought it to you on day 1!

🍎 macOS GA

Thanks to <@336691752645689345>, all major features of Lemonade, including OmniRouter and the embeddable lemond binary, are now available on macOS. Tell your friends to come join the party!

🧠 Canonical Model IDs

Users will never see the user. anymore and almost never the extra. prefix, even when using external software like Open WebUI, Pi, and Hermes. The new naming scheme is documented here.

Thank you <@372211157735309313> <@569452289195114502> <@589490993091313706> <@416634261961965579> <@477136961904312350> <@104003508562309120> <@169371491027255296> <@323176231098908672> for all the feedback that helped us get to a good design here!

🌍 Global and Per-Model Args

Before: if you had per-model llama.cpp (or whisper, or vllm, etc.) args they would overwrite your global args. Now they are merged. You can opt-out on a per-model basis.

Thank you <@569452289195114502> <@501720089113133057> <@375216812511723530> <@823901147123154975> <@1440101722365362440> <@477136961904312350> <@211398703238348800> for all the work and feedback on this and future improvements in this area!

Additional Improvements

  • Upgrade to FLM v0.9.42 and fix transcription GUI bugs by me, <@1331063909783830589>, and <@569452289195114502>
  • Easily copy canonical model IDs from the Model Options GUI by <@1440101722365362440>
  • vLLM and lemonade launch refinements by <@477136961904312350> and SorenDreano
  • Bug fixes by <@1325173172244578427>, <@693745725884071978>, and Gunnar

Breaking changes

  • room-preview has been deprecated because TheRock is now stable! You will be auto-migrated from rocm-preview to rocm-stable
  • audio has been renamed to transcription in the /v1//health endpoint
  • If you have *.args set in the global config and *_args set for a specific model, these options will be merged at load time. Previously, per-model args would completely replace global args. You can opt-out on a per-model basis by setting merge_args=False in a load request and/or saved model options.

Special should out to <@1440101722365362440> who gave a lot of great advice and feedback this cycle. Appreciate you!

See the full release notes

Clone this wiki locally