lemonade

Release candidate of PR 3930 (Lemonade Nexus) from branch commit ffd40a2, built by CI run 42558423737, with the Nexus SDK at lemonade-nexus 229726d.

This is a pre-release for testing the PR. It is not a Lemonade release.

New in this build:

• The device advertises its LAN addresses fastest link first and one per subnet, so peers try a 25 Gbit link before a 1 Gbit link and both before Wi-Fi.
• Together with Nexus server release v2026.11.09, two devices on one LAN whose sessions sit on different Tier 1 servers now see each other's LAN addresses and connect directly instead of through a relay.

From the previous builds: host-grouped chat model picker with Loaded/Downloaded marks, a mesh peer's load request keeps a loaded model, "Chat runs on" in the account menu, the server info dialog, the mesh port keeps its certificate across restarts, and Enable Mesh accepts the device's current account after a deletion or key replacement.

Headline

⚠️ These notes are AI generated and will be revised by a human ⚠️

• Benchmark results can be submitted directly from lemonade bench and shared via claim link at https://bench.lemonade-server.ai.
• Zero-shot classification is available via POST /v1/classify with a labels parameter, and as a zero_shot condition type in router policies.
• TheNoise image generation gains a CUDA backend for Linux x86_64 alongside existing ROCm support.
• New container backends for AMD Strix Halo (gfx1151) — including ds4:rocm, rocmfpx:rocm, gufo:rocm, and llamacpp:halobox — bring GPU-accelerated inference to the platform.
• POST /v1/systemone forwards typed questions to a decision model, with support for router-classified choices and SystemOne zero-shot classification.

Breaking Changes

⚠️ These notes are AI generated and will be revised by a human ⚠️

• The /models API response no longer includes image_defaults or audio_defaults; their values are now under recipe_options, so external clients reading the old fields will find them absent.
• The /models response includes a user_configurable boolean flag on backend descriptor options; consumers that assumed all listed options are editable may show technical knobs like soft_pixel_cap, pcm_sample_rate, and pcm_channels to users.
• FLM v1.0.11 (bumped from v1.0.7) returns max_tokens exactly as requested rather than max_tokens+1, so chat completions that relied on the extra token will now stop one token early.
• FLM v1.0.11 streams token usage counts only when stream_options.include_usage is set; without it the server's /api/v1/stats endpoint records no token counts or throughput for FLM streams.
• FLM v1.0.11 returns 400 for previously-accepted invalid fields including invalid tool_choice, forced tool calls on small Qwen3 models (up to 1.7B), conversation history on *-flash models, and requests with no messages or prompt.

Lemonade Server

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

Embe

Headline

⚠️ These notes are AI generated and will be revised by a human ⚠️

• AMD Strix Halo gets dedicated container backends for llama.cpp, ds4, gufo, rocmfpx, and halogen models.
• SystemOne decision models accept typed questions via POST /v1/systemone, with a new router classifier type.
• Zero-shot classification arrives at POST /v1/classify with a labels parameter and zero-shot classifier condition.
• TheNoise image generation gains a CUDA backend for Linux x86_64 alongside existing ROCm support.

Breaking Changes

⚠️ These notes are AI generated and will be revised by a human ⚠️

• The /models response no longer includes image_defaults or audio_defaults keys; their values moved under recipe_options, so external clients reading the old fields will find them absent.
• The /models response now includes a user_configurable boolean flag on backend descriptor options; consumers that assumed all listed options are editable may show technical knobs like soft_pixel_cap, pcm_sample_rate, and pcm_channels to users.

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.

Headline

⚠️ These notes are AI generated and will be revised by a human ⚠️

• The Lemonade app is redesigned with six workspaces, a visual router policy editor, and multi-provider model search.
• Five K2-Horizon and Flint-GGUF models added to the model registry.
• Models that previously failed to load on 128 GB Strix Halo systems now start reliably with improved shared-memory management.
• Model and backend update progress is visible directly in the app with server-owned sync jobs.

Breaking Changes

⚠️ These notes are AI generated and will be revised by a human ⚠️

• The resolved_ctx_size field on /v1/models/{model}/options may now be null when automatic context sizing cannot produce a valid context — clients must handle null instead of assuming an integer.
• GET requests to POST-only paths (slots/{id}, jobs/{id}/pause, jobs/{id}/interrupt, jobs/{id}/resume, models/register) return 405 instead of 404 — clients relying on 404 for these paths will need to check for 405.
• A request missing the model field now raises a consistent "Missing model field in request" error across all model routes (chat/completions, completions, embeddings, responses, rerank) — previous inconsistent error messages may no longer appear.
• Form-encoded requests with invalid field types (such as a non-numeric temperature on transcription) now return 400 instead of 500 — clients catching 500 for this condition should handle 400.
• Streaming responses carry a cancel token that stops the server from sending data after a TCP disconnect.

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.