Much of the work of the unlimited plan comes from making the most of the GPUs we have, using optimal settings for spec decoding.
This qwen 3.8 27B is a marvel, and at this speed, and without limits of use, it is greatly appreciated
Free: $20/month of AI agent usage.
No card. Perch AI Starter: web, desktop, CLI. → DeepSeek V4.1 Flash + GLM 5.3 Flash → Qwen 3.8 Max free until Nov 20 → Training on by default, opt out: Settings → Privacy Verdict: solid. Operator unnamed, keep secrets out.
New clean run with Tenaz v0.6.0
Same Raycaster Shooter challenge (prompt by ddalcu), Qwen 3.8 Flash Next on Sushi 1.5.1 with DeepSeek Harness, all local on an M5 Pro with 64 GB. One config from start to finish: no pauses, no hand edits. Compared with the first run: ⏱️ 19h20m →
Just ran Qwen 3.8 Flash Next Coder (pruned 125B model) on kaggle's free 2x T4 with Strata.
Generated this SVG at 28 tok/s.
You have reached the end of the archive
All of qwen38