Qwen-3.8-flash-next, after a lot of trouble, it finally worked.
His memory is short, so he's not very fast, but I get the impression that his thinking is very solid. The vision analysis is also deep. Waiting for vision completion from 1:00 to 3:00.
I use qwen-3.8-27B + Kokoro 82M on a RTX 6000 and I forced it to make “Just a Minute” explainer of this.
Did pretty good job
I ran the same todo-app prompt through the same agent harness a 3rd time.
Same model, same prompt. Only the process changed — and it changed everything. Run 2: 3h17m😲 (85 minutes of it burned by one subagent chasing a bug it introduced). Run 3: 43 minutes. 102 model calls,
This is wild, never thought I'd be running a 27B-parameter model on my phone.
Running Alibaba_Qwen's Qwen 3.8 27B (1-bit) on my iPhone 17 Pro. gonna share the benchmarks very soon. Stay tuned! You can try running it on your own phone using the RunAnywhereAI apps available on
You have reached the end of the archive
All of qwen38