Same speed. Better quality.
I've been daily driving the Qwen base model ported to MLX. So far this feels as fast as the 32B MoE and better quality than the 27B dense model. I haven't used many fine-tuned models so far, but this is the first that seems to work really well for agentic coding. Thanks to everyone that made it possible.
@TheRealMDT — really appreciate this. Hearing it
works for agentic coding is exactly what we designed it
for.
Since you're getting great results with the 3.6 MoE,
check out our brand new release — just shipped this
weekend:
Qwen 3.8-27B IQ-Injected — Unified Frontier Model
1,159,426 IQ-injected examples from 8 frontier labs.
The model thinks before it answers — reasoning baked
into weights. 93% benchmark across coding, security,
finance. Works with OpenClaw + Hermes out of the box.
Would love to hear how it compares for your agentic
coding workflow. Thanks for the feedback — it keeps
us shipping.
Excellent model, can't wait to run benchmarks on it :)
-G