Qwen3.6-35B-A3B-UD-MLX-4bit was surrprised me, running on mac mini m4 pro (48gb ram), with omlx backend + openclaude cli (also tested with pi).
Tested on my real work codebase, it's fully run with a bug fix planning/feature planning. but crashed on tools_use.
and just performed around 3xx PP/ 30~30 tok/s.
@ollama@ErRahul337 hope ollama can make owned coding cli, I am stick around claude code/pi/opencode for my real life working, but I hate always needs to patch config files for all of them, just only to make them works with ollama/custom provider/official provider.
Running Qwen3.6-35B-A3B-UD-MLX-4bit around 30~34 tok/s locally on single mac mini m4 pro (48gb ram onboard) with oMLX backend (enable TurboQuant option) and claude code.
in my opinion, it obviously works and not so bad.
@ClaudeDevs here is another problem I'm encountering is that I've set it to "auto effort", but Claude Code automatically sets it to the "max effort" level.
Spectra 0.2.4 is out. It's a macOS terminal built on libghostty.
- Fixed issue of Send Copy To was silently skipping same-named folders.
- Fixed issue of copying files to SMB/NAS shares failed with "Permission denied".
https://t.co/5ghMAT25Au
Shipped my side project: Spectra — a macOS terminal that *remembers* your setup.
Splits, tabs, directories → save once as a workspace.
Next time: one click, back to flow.
No more morning re-arranging.
https://t.co/5ghMAT25Au