0.16.8 — Flint support and concurrent short chats

Serve the Flint 40B mixture-of-experts checkpoint with tool calls and reasoning configured automatically. Concurrent short chats on MTP models decode together instead of queueing, and running servers surface update notices without a restart.

• Flint. rapid-mlx serve flint-40b-a2.5b-4bit serves the official INT4 checkpoint of a 40B-total, ~2.5B-active mixture-of-experts model. Tool-call and reasoning parsing are configured automatically for the alias. The validated checkpoint revision is pinned, so downloads, config and tokenizer reads, cache checks and loading use that revision; an unavailable pin fails instead of silently selecting another revision. Measured on an M5 Max 36 GB: ~1K context reached ~131 tok/s decode and 0.35–0.38 s to first token at about 24 GB peak memory; 48 GB or more unified memory is recommended. Speculative decoding and MTP stay off, and int8 KV cache is refused because quantized attention does not support the model's learned attention sinks. (#5079, #5086)
• Concurrent short chats. Four concurrent short MTP chats now decode together on M3 Ultra with qwen3.8-27b-4bit, raising aggregate throughput from 37.6 to 73.8 tok/s. Per-request time to first token changed from 0.6/4.0/7.3/10.8 s to about 1.4 s each. Single-stream speed is unchanged, and memory admission limits still apply. (#5085)
• DeepSeek V4.1 verification. Speculative verification chunks of 2–8 tokens use fused attention. On M3 Ultra 256 GB, measured target verification took about 6 ms less at a short prompt and 6–8 ms less at ~7K context; these are verification timings, not end-to-end decode gains, and single-row decode keeps its existing path. (#5043)
• Desktop. Model loading appears inside the chat composer, so startup status stays where you write your message. (#5056)
• Server updates. Text and audio servers re-check for releases every 24 hours and print a notice once per newly seen newer version. GET /health exposes latest_version and update_available; existing version-check opt-outs remain respected. (#5061)
• Faster first reply for coding-agent profiles. The optimized private launch enables only shell commands, file reading, writing and editing, and file and text search, reducing the tool schemas sent with the first request. Additional built-in tools remain available through manual setup. (#5084)