[hsliuustc0106]

What

• Collapsed the duplicated Supported-Models table into one matrix — feat(tts): VoxCPM2 serving (#71) + vendored compiled per-patch loop (#79) #76 (VoxCPM2) appended a second copy instead of editing the first, and both were rendering on main.
• ASR row: 🚧 planned → ✅ served — /v1/audio/transcriptions (#104); streaming + prefix cache pending (#68). Examples cell now Qwen3-ASR (0.6B/1.7B, 4-bit; decoder family) — the old "Whisper, Qwen3-ASR, Voxtral, …" list was aspirational; asr/capabilities.py boot-rejects every family except qwen3_asr today.
• TTS row: added MOSS-TTS-Nano-100M (tts: add MOSS-TTS-Nano-100M backend branch, config and loader (#73), … #110).
• Latest News: bullets for ASR served (feat(asr): /v1/audio/transcriptions adapter over qwen3_asr (#68 tasks 1–2) #104, ASR decoder-style first: /v1/audio/transcriptions on qwen3_asr — prompt-prefix cache + TPOT/ITL doctrine #68 tasks 1–3) and VoxCPM2 3.5× + MOSS-Nano (perf(tts): VoxCPM2 3.5× — timestep knee + quality-gated 8/4-bit DiT/encoder quantization (#88) #98/VoxCPM2 speedup phase 2: timestep knee + DiT/encoder quantization (the bandwidth floor) #88, tts: add MOSS-TTS-Nano-100M backend branch, config and loader (#73), … #110/TTS model: MOSS-TTS-Nano-100M (tier 2) — tiny single-stage AR, cloning-only #73).
• Getting Started / footprint: pip install -e '.[asr]' line + note that [asr] pins the same mlx-audio as [tts] and adds only python-multipart.
• Options: --omni text now matches the actual CLI help (a Qwen3-TTS, VoxCPM2, or MOSS Nano checkpoint serves /v1/audio/*); --asr-model listed.
• Design Notes: "speech in (ASR via mlx-audio stt) is planned" → served via the ASR endpoint, streaming/prefix-cache noted as pending.

Why

The README lagged main by four merges: #104's merge added the API-table row and usage snippet but never updated the model matrix, and #98/#110 landed with no news entry. Claims verified against code at 210cc4e: --asr-model in __main__.py, [asr] extra in pyproject.toml, the capabilities matrix serving only qwen3_asr, and moss_tts_nano requiring --ref-audio.

Docs-only — no .py touched (ruff vacuous). python -m unittest discover -s tests → 233 tests OK, 25 skipped (weight-gated).

Notes for reviewers

• Overlap: feat: diffusion engine + Z-Image-Turbo on /v1/images/generations (#91, #99) #108/feat(image): mflux seam + Qwen-Image-2.1 serving — /v1/images/generations, Viggle turbo fast lane (#91, #101) #109 (image PRs) still carry the old duplicated table plus their new image rows — they'll want a small rebase over this dedupe; their image rows are unaffected. docs(speech): quiet-session per-checkpoint table; clone TTFA corrected (0.6B = small-memory pick) #113 touches only docs/speech.md, no overlap.
• The Diffusion row stays 🚧 planned on main — image serving is unmerged (feat: diffusion engine + Z-Image-Turbo on /v1/images/generations (#91, #99) #108/feat(image): mflux seam + Qwen-Image-2.1 serving — /v1/images/generations, Viggle turbo fast lane (#91, #101) #109 open).
• docs/speech.md has no MOSS-Nano section yet (owed tts: add MOSS-TTS-Nano-100M backend branch, config and loader (#73), … #110 follow-up); the TTS row links to the guide as-is.