Headline
- A new experimental
ds4 (DwarfStar) backend serves DeepSeek V4 Flash on AMD Strix Halo GPUs.
- The catalog adds the RPG-HaloTales-V2 chat model, Flux-2-Klein-4B/9B image models, and new OpenMOSS voice-design and sound-effect speech models.
- Cloud providers now support custom authentication headers and byte-for-byte Anthropic Messages passthrough via
wire_format, and lemonade cloud list --json reports provider readiness.
- Interrupted model downloads can be cancelled with Ctrl+C and resumed, capped with a new
download_rate_limit config key, and long connections stay alive through TCP keepalive and streaming heartbeats.
- The new
lemonade update-models command and global and per-model auto-update settings keep installed models current.
Breaking Changes
LemonadeServer.exe --port and --host are now ephemeral in-memory overrides and no longer persist to config.json; use lemonade config set port=<P> / host=<H> to persist them.
- Custom
*_args precedence changed: explicit request *_args no longer inherit model/architecture args, and merge_args=false now suppresses all inherited backend/machine args and overridable runtime defaults.
- The llama.cpp backend now defaults
--parallel to 1, removing implicit multi-slot concurrent batching; restore it with --llamacpp-args "--parallel N".
- Legacy telemetry attributes
openinference.session.id and openinference.user.id were removed in favor of the standard session.id and user.id keys; update downstream dashboards and queries.
- whisper.cpp transcription now defaults an omitted
language to auto (source-language auto-detection) instead of English; pass language=en to force English.
- Removed the undocumented HTTP routes
/v1/params, /v1/log-level, /api/v1/test, and /status; use /internal/set and /internal/config instead.
- Backend
latest version resolution now includes pre-releases for all backends.
Lemonade Server
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.
What's Changed
Thanks @Bekhouche, @GabrielReusRodriguez, @Geramy, @NineBallo, @abn, @anditherobot, @bitgamma, @blackdeathdrow, @bong-water-water-bong, @fl0rianr, @github-actions, @jeremyfowers, @jgmelber, @kenvandine, @meghsat, @original4422, @pwilkin, @ramkrishna2910, @sjjh, @soothill, @superm1, @zaneni6 for your awesome contributions to this release!
Click to expand changelog
- Add missing handler for lemonade:// URLs (Closes: #1144656) by @superm1 in #3195
- fix(GUI app): stop chat from bypassing collection.router routing by @meghsat in #3180
- Fix the developer setup page URLs by @superm1 in #3225
- fix(rocm): only stage amdhip64_7.dll when TheRock runtime is newer than System32 by @blackdeathdrow in #3217
- docs: add latest release and Lemonade AI YouTube videos to news feed by @jeremyfowers in #3202
- feat(cloud): configurable auth header name and prefix per provider by @jgmelber in #3222
- fix: cancel model download on Ctrl+C (SIGINT) via libcurl progress callback (#1385) by @bong-water-water-bong in #2462
- feat(server): record the backend launch command in /health by @anditherobot in #3229
- chore: rename app identifiers to ai.lemonadeserver namespace by @abn in #1978
- docs: add MCP Gateway to the API Spec nav by @jeremyfowers in #3242
- Update the contribution guide for H2 2026 by @jeremyfowers in #3246
- Add HaloTales-V2 and related models by @bitgamma in #3264
- Flm version update by @zaneni6 in #3266
- fix(server): treat --port and --host CLI flags as ephemeral overrides by @abn in #3249
- Move persistent Lemonade JSON state from
.cache to .config by @superm1 with @Copilot in #3028
- feat(server): hide backends with no runnable models on this host by @ramkrishna2910 in #3219
- feat(telemetry): configurable session ID ingestion and OpenInference tool-call capture by @abn in #3248
- feat(server): treat config.json as sparse user overrides with deep merging by @abn in #3268
- test(telemetry): prevent dangling RuntimeConfig pointer and worker race in telemetry helpers test by @abn in #3283
- fix(server): resolve custom args by scope by @fl0rianr in #3265
- feat(server): enable cross-platform tcp keepalive socket options by @abn in #3209
- ci: temporarily disable tts-openmoss integration tests by @Geramy in #3291
- test: fix dangling CLI temp path in runtime override test by @fl0rianr in #3288
- test: move multi-checkpoint download-state coverage to C++ by @fl0rianr in #3286
- test: move llamacpp system backend policy coverage to C++ by @fl0rianr in #3287
- test: migrate model registry label validation to C++ by @fl0rianr in #3285
- feat(server): add context_length to the models endpoints by @anditherobot in #3280
- feat(cli): add --json output to lemonade cloud list by @GabrielReusRodriguez in #3261
- feat(streaming_proxy): add sse comment heartbeats during prompt prefill by @abn in #3210
- test(server): assert CLI parse return status and remove unused header in runtime override test by @abn in #3294
- fix(router): handle transient GPU hangs and compute errors in lemond by @abn in #2652
- feat(sync): add administrative model synchronization and auto-update configuration by @abn in #2779
- Update stable-diffusion.cpp to master-827-97d2990 and CUDA master-827-e2329c3 by @github-actions[bot] in #3311
- [Stable Diffusion] Respect the user configurable "Model Options" by @NineBallo in #2845
- Handle llama.cpp new release scheme by @bitgamma in #3315
- Update faq.md: correcting list format by @sjjh in #3299
- fix(server): release large request memory by @soothill in #2873
- feat(server): size-filter streaming backends by working set, not full size by @ramkrishna2910 in #3328
- docs: document POST /v1/install/dry-run by @jeremyfowers in #3324
- feat(router): add min_total_chars / max_total_chars whole-conversation length conditions (#2958) by @meghsat in #3181
- docs: document POST /internal/simulate-vram-pressure by @jeremyfowers in #3325
- feat(routing): add expected_output_tokens signal to RouteContext by @Bekhouche in #3163
- docs: document POST /v1/routing/validate by @jeremyfowers in #3323
- Retire undocumented /params, /log-level, /api/v1/test, and /status routes by @jeremyfowers in #3322
- Add TheNoise-based upscalers by @bitgamma in #3327
- Require 2 reviewers for breaking changes by @jeremyfowers in #3348
- ci: retry apt-get on transient arm64 mirror 404s by @kenvandine in #3353
- fix(vllm): scale gpu-memory-utilization to free memory on shared-memory devices by @jeremyfowers in #3344
- test(server): eliminate timing flakiness in StreamingHeartbeatTest by @abn in #3338
- fix(openmoss): bump OpenMOSS TTS pin to v0.3.0 and re-enable CI by @jeremyfowers in #3335
- docs: add smart router subject area and two router maintainers by @ramkrishna2910 in #3351
- Bump project version from 11.7.0 to 11.8.0 by @jeremyfowers in #3354
- Update faq.md removed double tts chapter by @sjjh in #3300
- feat(cli): return per-model detail from status --json by @anditherobot in #3259
- fix(llamacpp): pin --parallel 1 so ctx_size is not oversubscribed by @jgmelber in #3277
- feat(config): add configurable download rate limit by @blackdeathdrow in #3281
- fix(whisper): auto-detect omitted transcription language by @original4422 in #3245
- Feat/openmoss 0.3.0 backport/server side by @pwilkin in #2863
- feat(cloud): relay Anthropic-wire-format providers through /v1/messages by @jgmelber in #3223
- feat(backends): ds4 (DwarfStar) backend for DeepSeek V4 Flash [experimental] by @ramkrishna2910 in #3047
New Contributors
Full Changelog: v11.7.0...v11.8.0
Windows installers are signed. Free code signing provided by SignPath.io, certificate by SignPath Foundation. See our Code Signing Policy.
