chore: release main

[github-actions]

🤖 I have created a release beep boop

higgs-bench: 1.0.3

Dependencies

higgs: 2.0.0

2.0.0 (2026-09-19)

⚠ BREAKING CHANGES

• higgs start is now config/profile-only, attach is a strict daemon dashboard, shellenv/exec fail fast on invalid or unreachable targets, and exact local model matches now take precedence over regex routes.

Features

• add --profile CLI flag, TUI routing tab, and config visibility (#45) (95dda05)
• add higgs exec -- <command> subcommand (#49) (b61dbf5)
• add higgs run -- <command> subcommand (#47) (72339d2)
• add name field to ModelConfig (#43) (54147e0)
• add qwen3.5/qwen3.6 turboquant stack (4f165ee)
• API routes — thinking mode streaming and tool call validation (5f7e62c)
• chat: streaming tool-call deltas + Qwen-friendly arg normalisation (#164) (8f01163)
• cli: higgs ui opens the desktop app (09a5b4c)
• config: first-class config surface for the MLA latent KV cache (#263) (cdb3c4e)
• desktop: Tauri dashboard app, request tracing, and richer server metrics (72d8809)
• desktop: Tauri dashboard app, request tracing, and richer server metrics (487716f)
• doctor: validate ServerSection fields (0250c42)
• doctor: validate ServerSection fields (554371f)
• engine: chunked prefill yield quantum for the batch serving loop (PR8 / ds4 P9) (#259) (247147e)
• engine: disk-backed KV prefix store for restart-resume (PR6 / ds4 P4) (#257) (0435d11)
• harden CLI, dashboard, routing, and MLX runtime (4dfc930)
• harden default security posture (1b41a5c)
• higgs: migrate from huggingface-cli to hf (#180) (aaecdb4)
• resize routing sidebar (913e015)
• server config — TurboQuant, thinking budget, chunked prefill settings (5b63e4c)
• serve: stream prefill progress (llama.cpp-compatible prompt_progress) (#184) (e7b2c8a)
• ship mlx.metallib alongside the higgs binary (#39) (deaa322)
• tune qwen3.6 thinking defaults (6969834)
• unified AI gateway with proxy routing and format translation (#38) (7c5668b)

Bug Fixes

• address CI lint and test failures (0d8e259)
• address coderabbit issues (6f10bc6)
• address follow-up review issues (5c857ce)
• address review comments (36c6484)
• address review follow-ups (d2b312e)
• backtick TcpListener in doc comment for clippy doc_markdown (94acea7)
• cargo fmt + restore higgs crate build on PR #74 (050af95)
• clarify metallib recovery hint (645dbb0)
• clear lint and review blockers (9c8b878)
• config: timeout validation and figment model merge (441884d)
• correct comment about signal handler behavior (81cbd71)
• default Qwen3.6 to non-thinking mode (#150) (372020e)
• desktop: address review findings and bundle the CLI in the app (bb9a5c3)
• doctor: bracket IPv6 host when binding port; add api_key/rate_limit positive-path tests (7177476)
• higgs: validate config writes, and make metrics see failed requests (#272) (cfe0ab0)
• improve Anthropic API compatibility (16ee50d)
• improve Anthropic API compatibility (ff7e96c)
• log warning instead of silently discarding signal handler error (6946974)
• make pre-push pass on dust stack (2695165)
• normalize auto_router.model to a name like routes do (1e741c5)
• normalize auto_router.model to a name like routes do (3b52db4)
• remove incorrect default_value on Option<u8> CLI arg — PR #75 (bf87bbd)
• resolve auto_router model by basename when config uses full path (#41) (7220fa2)
• restore higgs crate build on PR #75 — chat.rs usage stubs (f92d139)
• restore pattern matching fallback in force routing mode (c1bb2c3)
• restore pattern matching fallback in force routing mode (217249f)
• route first streamed token through incremental detokenizer (722218d)
• routes: close think tag on length-stopped reasoning in OpenAI path (5c2ea56)
• satisfy clippy duration lint (20c6aae)
• split long first doc paragraphs for clippy (c195312)
• stabilize dust stack CI (8e629e3)
• strip thinking tags from Anthropic route, fix force routing, add proxy usage tracking (89f8db2)
• strip thinking tags, fix force routing, add proxy usage metrics (0ffbabb)
• suffix float literals for new float_literal_f32_fallback lint (#230) (83029e4)
• support clippy across toolchains (abc316d)
• support mlx qwen3.6 smoke (7046849)
• type alias for build_router signature assertion (d084592)
• use 128+signal convention for signal-killed child exit codes (89b7ae0)

Performance Improvements

• sse: pre-serialize static chunk fields for streaming routes (056f4b4)
• sse: pre-serialize static chunk fields for streaming routes (d8e02ba)

higgs-engine: 2.0.0

2.0.0 (2026-09-19)

⚠ BREAKING CHANGES

• higgs start is now config/profile-only, attach is a strict daemon dashboard, shellenv/exec fail fast on invalid or unreachable targets, and exact local model matches now take precedence over regex routes.

Features

• add qwen3.5/qwen3.6 turboquant stack (4f165ee)
• bench: context-frontier benchmark with shared cache rollback helper (PR2 / ds4 P8) (#253) (4510835)
• bonsai-q1: packed engine scaffold with upstream MLX guard (#142) (fe43aab)
• bonsai: run Bonsai-Q1 bits=1 on vanilla MLX via JIT qmv_fast kernels (no mlx-rs fork) (#182) (b604754)
• chat: streaming tool-call deltas + Qwen-friendly arg normalisation (#164) (8f01163)
• enable MiniCPM5-1B (explicit head_dim + special-token decode + tojson kwarg + tool parser) (#176) (05464e2)
• engine: chunked prefill yield quantum for the batch serving loop (PR8 / ds4 P9) (#259) (247147e)
• engine: disk-backed KV prefix store for restart-resume (PR6 / ds4 P4) (#257) (0435d11)
• harden CLI, dashboard, routing, and MLX runtime (4dfc930)
• harden default security posture (1b41a5c)
• inference engine — chunked prefill, prefix cache, thinking budget, MTP (4fa0971)
• models: add Gemma 3 and Gemma 4 inference support (4f301fc)
• models: MLA latent KV cache with weight absorption for DeepSeek-V2 (PR3 / ds4 P1) (#254) (482d2ec)
• mtp: release speculative decoding optimizations (0e5e458)
• mtp: release speculative decoding optimizations (1d15599)
• Qwen3.5 model architecture, TurboQuant KV cache, and model-level optimizations (1514737)
• qwen35: Qwen3.6 MoE MTP speculative decode + checkpoint memory-safety fix (#183) (4989955)
• serve: stream prefill progress (llama.cpp-compatible prompt_progress) (#184) (e7b2c8a)
• thinking mode support — reasoning parser and chat template (a3cc52e)
• tune qwen3.6 thinking defaults (6969834)

Bug Fixes

• address coderabbit issues (6f10bc6)
• address follow-up review issues (5c857ce)
• address remaining draft review issues (4396e88)
• address review items across models and engine — PR #74 (8863e2d)
• cargo fmt + restore higgs crate build on PR #74 (050af95)
• clear lint and review blockers (9c8b878)
• deps: migrate to sha2 0.11 (#273) (eb817f9)
• doc_markdown lints in detok test comments (deefbe3)
• engine,models: cache config propagation, ndim guards, FSM ordering (37d2ddd)
• engine,models: clippy clean lib targets on PR #74 (51→0 errors) (ebf3192)
• engine: clippy clean 6 files on PR #74 (85→51 errors) (7dc7370)
• engine: prefix cache full-hit, scheduler leak, gather bounds check (3e35b3f)
• make pre-push pass on dust stack (2695165)
• rename first-token detok bindings to avoid shadowing (8d16696)
• route first streamed token through incremental detokenizer (722218d)
• split long first doc paragraphs for clippy (c195312)
• stabilize dust stack CI (8e629e3)

Performance Improvements

• incremental detokenization for streaming generation (a949e50)

higgs-models: 2.0.0

2.0.0 (2026-09-19)

⚠ BREAKING CHANGES

• higgs start is now config/profile-only, attach is a strict daemon dashboard, shellenv/exec fail fast on invalid or unreachable targets, and exact local model matches now take precedence over regex routes.

Features

• add qwen3.5/qwen3.6 turboquant stack (4f165ee)
• bench: context-frontier benchmark with shared cache rollback helper (PR2 / ds4 P8) (#253) (4510835)
• bonsai-q1: packed engine scaffold with upstream MLX guard (#142) (fe43aab)
• bonsai: run Bonsai-Q1 bits=1 on vanilla MLX via JIT qmv_fast kernels (no mlx-rs fork) (#182) (b604754)
• cache: AnyCache::trim_by dispatcher for spec-decode rollback (#143) (229c111)
• config: first-class config surface for the MLA latent KV cache (#263) (cdb3c4e)
• enable MiniCPM5-1B (explicit head_dim + special-token decode + tojson kwarg + tool parser) (#176) (05464e2)
• engine: disk-backed KV prefix store for restart-resume (PR6 / ds4 P4) (#257) (0435d11)
• harden CLI, dashboard, routing, and MLX runtime (4dfc930)
• models: add Gemma 3 and Gemma 4 inference support (4f301fc)
• models: MLA latent KV cache with weight absorption for DeepSeek-V2 (PR3 / ds4 P1) (#254) (482d2ec)
• models: per-tensor quantization settings with dense-mode loading (PR9 / ds4 P2 loader half) (#260) (92277ca)
• quantize: MoE calibration tooling + loader fixes; asymmetric-quality claim negative (#264) (9c40cce)
• qwen3_next: mixed-bit Qwen3.5 GDN BA loading fallback (#148) (cc18616)
• Qwen3.5 model architecture, TurboQuant KV cache, and model-level optimizations (1514737)
• qwen35: Qwen3.6 MoE MTP speculative decode + checkpoint memory-safety fix (#183) (4989955)
• serve: stream prefill progress (llama.cpp-compatible prompt_progress) (#184) (e7b2c8a)

Bug Fixes

• address coderabbit issues (6f10bc6)
• address review items across models and engine — PR #74 (8863e2d)
• align turboquant tests with rand 0.10 (d0617c4)
• bonsai: apply causal mask during prefill (#203) (ecd78e5)
• cargo fmt + restore higgs crate build on PR #74 (050af95)
• clear lint and review blockers (9c8b878)
• deps: update rust crate safetensors to 0.7 (#151) (a177ef6)
• deps: update rust crate safetensors to 0.8 (#201) (27d20e2)
• engine,models: cache config propagation, ndim guards, FSM ordering (37d2ddd)
• engine,models: clippy clean lib targets on PR #74 (51→0 errors) (ebf3192)
• gate rng import to tests (1161eaa)
• honor qwen3 next gate quantization (00958e1)
• make pre-push pass on dust stack (2695165)
• models: clippy clean for higgs-models on PR #74 (part 2/2) (1b4470b)
• models: clippy clean in 5/7 files on PR #74 (part 1/2) (af57c32)
• resolve new clippy lints from stable toolchain update (#286) (8fb011c)
• restore rng import and checkout pin (8771fa0)
• satisfy clippy on qwen3.6 tests (ab2920a)
• stabilize dust stack CI (8e629e3)
• suffix float literals for new float_literal_f32_fallback lint (#230) (83029e4)
• support mlx qwen3.6 smoke (7046849)

Performance Improvements

• dtype: preserve fp16 through scalar multiply in deepseek_v2 + siglip (5dfbd15)
• dtype: preserve fp16 through scalar multiply in deepseek_v2 + siglip (90aec29)
• models: opt-in fused MoE gate+up — 3→2 expert matmuls per layer (#141) (60d7cb4)
• sampling: partial-sort path for small top-k (67a5a06)
• sampling: partial-sort path for small top-k (bab4519)

---

This PR was generated with Release Please. See documentation.