commits

SHA	Message	Author	Date
e281f0bb	feat(muse_glimmer): land M0 and MXFP4 conversion (#117) Co-authored-by: Claude Opus 5 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	16 days ago
f8d27417	docs: group supported models by type	LongYinan <l**m@g**m>	18 days ago
04d68502	docs: rewrite README	LongYinan <l**m@g**m>	18 days ago
37f1d68e	Complete continuous batching through Stage 2 (#116) Co-authored-by: Claude Fable 5 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	18 days ago
2d1fe60e	docs(experiments): bank the 2026-08 NAX research arc as a record-only worklog Co-authored-by: Claude Fable 5 <n**y@a**m>	LongYinan <l**m@g**m>	22 days ago
6eb2b040	feat(mtp): acceptance-gated speculative decoding (fall back to AR on weak dra...	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	24 days ago
23ee7d68	perf(qwen3_5,lfm2): cache per-layer kinds; docs: align with deleted compiled-...	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	24 days ago
34c53f40	feat: add Qwen3-ASR conversion, streaming, and meeting capture (#112)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	27 days ago
237ab8a8	0.0.10	LongYinan <l**m@g**m>	29 days ago
8ec1d5b8	perf: optimize quantized inference fast paths (#113)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	29 days ago
96d56562	perf(gemma4): accelerate long-context decode (#111)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	30 days ago
5447ded3	fix(desktop): surface control panel on launch (#110)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
44d21532	fix(release): keep dashboard private	LongYinan <l**m@g**m>	about 1 month ago
33750f79	0.0.9	LongYinan <l**m@g**m>	about 1 month ago
082b7f7c	fix(ci): invoke desktop packager through root	LongYinan <l**m@g**m>	about 1 month ago
38e189d2	0.0.9	LongYinan <l**m@g**m>	about 1 month ago
4fef3b33	fix(chat): preserve template continuation cache reuse (#109)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
f19b14db	fix(chat): make model templates authoritative for all continuations (#108)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
91f232bd	feat(cli): revision-aware mlx download model — detect upstream updates, sync ... Co-authored-by: Claude Fable 5 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
e6598c6b	fix(agent): recover Gemma 4 tools, images, and restart cache (#106)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
400b12b6	feat(desktop): mlx-node.app — macOS menubar app with a supervised inference s... Co-authored-by: Claude Opus 5 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
5e67047c	feat: add Gemma4 Unsloth tensor-class recipes (#105)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
573ea23c	fix(test): make the two e2e gates that broke main deterministic (#102) Co-authored-by: Claude Opus 5 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
d0b608fa	feat: mlx dashboard — local web UI for models, sessions, metrics, and paged p... Co-authored-by: Claude Fable 5 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
b2ec89e4	docs: add convert/quantize internals reference Co-authored-by: Claude Opus 5 (1M context) <n**y@a**m>	LongYinan <l**m@g**m>	about 1 month ago
e59d78e7	docs: cleanup the CLAUDE.md	LongYinan <l**m@g**m>	about 1 month ago
b89b84cb	feat(quant): native ggml K-quants (Q6_K/Q4_K/Q5_K) — bit-exact, on the NAX te... Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
0ebeaa57	0.0.8	LongYinan <l**m@g**m>	about 1 month ago
f19d369d	feat(convert): add Unsloth MXFP and DGX maps	LongYinan <l**m@g**m>	about 1 month ago
b9260e77	feat(convert): allow fixed Unsloth maps without imatrix (#99)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
d53f2c0b	fix(agent): contextualize cached prefill metrics (#98)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
72603428	feat(agent): coordinate shared local inference (#97)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 1 month ago
5602f12d	feat(convert): add official Unsloth MXFP and NVFP4 maps for Qwen3.5/3.6 (#96)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
da84f23b	feat(agent): mlx agent — first all-in-one local coding agent on pi (#95) Co-authored-by: Claude Fable 5 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
51383108	refactor(engine): per-turn execution planning (TurnPlan/ExecutionPlan) + fail... Co-authored-by: Claude Fable 5 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
6417815b	fix(ci): drop redundant reference in format! arg (clippy 0.1.97 gate) (#93) Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
59ba19ab	chore: gitignore /Model-Optimizer local clone Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@g**m>	about 2 months ago
31412286	feat: NVIDIA modelopt recipe for mlx — --q-recipe nvidia (MXFP4 weights) + ml... Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
31f3502b	fix(convert): keep recipe-pinned o_proj/out_proj/in_proj_a/b 8-bit affine und... Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
84f2fc5b	feat(gemma4): native MTP via Google's official assistant draft checkpoints (#88) Co-authored-by: Claude Fable 5 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
b33c182e	fix(convert): AWQ Group A folds into pre_feedforward_layernorm on sandwich-no...	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
c84764fe	feat(gemma4): DSpark speculative decoding (DeepSeek DeepSpec external draft) ... Co-authored-by: Claude Fable 5 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
4e350f7a	fix(qwen3.5): allow non-affine (nvfp4/mxfp8) quantized lm_head + MTP fc via L... Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
c1bb68a1	test(qwen3.5-MTP): de-flake mid-cycle-cancel heal e2e + deterministic heal co... Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
4c65f808	perf: inference perf polish + quality sweep (qwen3.5 RoPE axis fix, MLX NAX-f... Co-authored-by: Claude Sonnet 5 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	about 2 months ago
f65104aa	fix: align repetition handling with vLLM (cutoff off by default) (#82) Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
73f4a89e	feat(server): honor stop_sequences in /v1/messages + loosen Qwen repetition c... Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
855b91cc	fix(qwen3.5-VL): correct image inference (interleaved M-RoPE, patch-embed, co... Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
ed672d16	test(losses): de-flake test_cross_entropy_qwen3_vocab (batch 2 → 64) (#79) Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
409a7b5b	fix(convert): apply imatrix AWQ on VLM checkpoints + compensate GDN in_proj_a... Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
4a474a9b	feat(gemma4): load tied lm_head as packed-quantized (8-bit affine) (#77) Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
4be27cf5	feat(convert): quantize gemma4_unified (12B) — recognize + keep vision_embedd... Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
5aa5759f	chore(deps): refresh JS + Rust dependencies (#75)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
441afcf1	feat(gemma4): Gemma 4 12B unified (gemma4_unified) — encoder-free text + visi... Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
6deb4508	feat(gemma4): Gemma 4 E2B QAT (wNa8o8) mobile-transformers support (text + vi... Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
ee88b92b	Chat-engine refactor: model-neutral engine (P1–P8) + honor generation_config.... Co-authored-by: Claude Fable 5 <n**y@am>, p6 <p6@l**l>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
74f981e8	perf(qwen3.5): CUDA GDN chunked-ops prefill (~1.3–1.75× TTFT, all Qwen3.6 cla... Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	2 months ago
87a59fca	feat: aarch64-unknown-linux-gnu (NVIDIA GB10 / DGX Spark) CUDA backend suppor... Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	3 months ago
fba240b8	perf(qwen3.5): GDN per-step prefill default + cold-mmap load prewarm (all fam... Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	3 months ago
36e54bec	perf(lfm2): quantized flat-default (~1.84×) + compiled flat/paged decode (~2.... Co-authored-by: Claude Opus 4.8 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	3 months ago
a90def75	feat(qwen3.5): MTP speculative decoding for Qwen3.5/3.6 dense + MoE (#65) Co-authored-by: Claude Opus 4.7 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	3 months ago
3360e4f8	perf(lfm2): compiled paged decode (~2× / ~1.5×) + paged-prefill last-token sl... Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	3 months ago
def5a1aa	feat: LFM2.5 MoE (lfm2_moe / LFM2.5-8B-A1B) inference support (#64) Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	3 months ago
35559218	fix(cli): make HF token download-wide; support `mlx download --set-token` Co-authored-by: Claude Opus 4.8 <n**y@a**m>	LongYinan <l**m@g**m>	3 months ago
cf654d43	fix(convert): unblock 250GB+ MoE conversions (CPU stream + drain on write) (#63) Co-authored-by: Claude Opus 4.7 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	3 months ago
95d46dbd	perf(qwen3_5): -16.4% prefill — FlashQLA-direction autoresearch arc + E53 sca... Co-authored-by: Claude Opus 4.7 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
434bfbfe	feat(quant): plumb MXFP4/MXFP8/NVFP4 through compiled MoE path; --q-recipe + ...	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
025e5af8	feat: openai/privacy-filter port — PII detection & redaction (#60) Co-authored-by: Claude Opus 4.7 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
7af2a538	fix: anthropic /v1/messages parsing, observability, and is_error (#58) Co-authored-by: Claude Opus 4.7 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
7faf106d	chore: read HUGGINGFACE_TOKEN env var as fallback for mlx download (#59) Co-authored-by: Claude Opus 4.7 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
df44f346	docs: update CLAUDE.md	LongYinan <l**m@g**m>	4 months ago
1fc7d166	chore: remove dead code across inference control flow (#57) Co-authored-by: Claude Opus 4.7 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
95dc24fa	perf: optimize inference control flow tracing and cache reuse (#56)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
df663fdf	feat: paged attention (#55) Co-authored-by: Claude Opus 4.7 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
7daf69de	feat: mlx launch claude, server-side tools, tier-2 prefix cache (#52) Co-authored-by: Claude Opus 4.7 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
144653d3	fix: make serve runnable (#51) Co-authored-by: Claude Opus 4.7 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
268d046d	fix: accept output_text, reasoning, and input_image in /v1/responses mapper (... Co-authored-by: Claude Opus 4.7 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	4 months ago
bfdeef52	0.0.7	LongYinan <l**m@g**m>	5 months ago
a3974fa8	chore: update deps	LongYinan <l**m@g**m>	5 months ago
0e095f04	refactor: session API (#48) Co-authored-by: Claude Opus 4.6 <n**y@am> Signed-off-by: LongYinan <gb@l**e>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
17a4fbcc	feat: Adds chat_stream support for gemma4 (#46)	Ivo von Putzer Reibegg <i**r@gm> Committed by: GitHub <ny@g**m>	5 months ago
1163c140	feat: add LFM2.5-1.2B-Thinking model (LiquidAI hybrid conv+attention) (#47) Co-authored-by: Claude Opus 4.6 <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
b04c95cf	feat: OpenAI responses api and Anthropic Messages API (#45)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
21bf41cb	feat: support Gemma4 models (#43) Co-authored-by: Claude Opus 4.6 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
aaf270c8	docs: add await to model.chat call in README example (#44) Signed-off-by: Ivo von Putzer Reibegg <i**r@g**m>	Ivo von Putzer Reibegg <i**r@gm> Committed by: GitHub <ny@g**m>	5 months ago
dd3df1a0	feat: add reasoning control flow for Qwen3.5 models (#41) Co-authored-by: Claude Opus 4.6 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
11f8caa5	feat: add Harrier embedding model (microsoft/harrier-oss-v1) (#40) Co-authored-by: Claude Opus 4.6 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
2024eefd	refactor: extract shared persistence and chat helpers (#39)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
1e1c5fcf	feat: add Qianfan-OCR (#38) Co-authored-by: Claude Opus 4.6 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
b431b9b3	fix: cleanup Codex review issues (#37) Co-authored-by: Claude Opus 4.6 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
1c1ae237	feat: stop generation on </tool_call> token when tools are active (#36) Co-authored-by: Claude Opus 4.6 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
999b9dbb	feat: add presence_penalty and frequency_penalty support (#35)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
6a8cef06	0.0.6	LongYinan <l**m@g**m>	5 months ago
3007f5b3	fix: chunked prefill + causal SDPA to prevent OOM on long contexts (#34)	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
00193576	0.0.5	LongYinan <l**m@g**m>	5 months ago
a305cb8b	feat: prompt cache reuse (#33) Co-authored-by: Claude Opus 4.6 (1M context) <n**y@a**m>	LongYinan <l**m@gm> Committed by: GitHub <ny@g**m>	5 months ago
2135ea7b	ci: enable incremental	LongYinan <l**m@g**m>	5 months ago
6d0c7888	0.0.4	LongYinan <l**m@g**m>	5 months ago
5853d9b7	chore: add internal tools	LongYinan <l**m@g**m>	5 months ago
7784fa85	chore: update fmt config and typecheck options	LongYinan <l**m@g**m>	5 months ago

← Back to repository

ecosyste.ms

Data

Tools

Indexes

Applications

Experiments

Commits

GitHub / mlx-node/mlx-node / commits