Repository navigation
Dyno Lab 0.6.16
Dyno Lab 0.6.16 shows what a model server is doing: Performance → Logs follows each request as it reads the prompt and generates, so a long first prompt from Claude Code no longer looks like a hang. The DFlash2 draft model can now be found by name in Discover, and the app says "speculative decoding" and "draft model".
Models and Discover
- Find the draft model by name. Searching Discover for
z-lab/Qwen3.8-27B-DFlash2, "dflash", "draft" or "speculative" now shows it, with what it does and its download state. Hugging Face tags it as a Transformers model, so the hub search never returned it. Downloaded has a Draft models section, where you can delete it. - Plain words. The button is now Speculative decoding (then Speculative decoding ready) instead of "Get drafter", and the app,
dyno launchanddyno servesay "draft model". "Good with Dyno" names the draft model under Qwen3.8-27B. - Full models with "DFlash2" in their name stay listed. Dyno now tells a draft model from its
config.json, asdyno launchdoes, so a download such asGLM-5.3-Flash-…-DFlash2no longer disappears from Downloaded.
Performance
- Logs. Performance has a new Logs source: what a model server is doing right now (waiting, reading the prompt with tokens read, speed and time left, or generating) and its full log, live. Servers started in Models now keep their log in
~/.mlx-dyno/logs/serve-<port>.log;dyno launchservers are listed too. - The server says what it's doing.
dyno servewrites a line when a request arrives, starts reading its prompt (and how much came from the prompt cache), gets its first token and finishes, plus a progress line every 10 s while reading a long prompt and a warning when no token has come for 30 s. With a draft model or MTP head this progress was invisible before./statsreports the same underlive.requests.