Unsloth ships Qwen3.8-Flash-Next Dynamic v3.1 GGUFs with 10-15% accuracy gains and engram switching
danielhanchen · reddit · 2026-11-05
Unsloth released Dynamic v3.1 GGUF quantizations for Qwen3.8-Flash-Next, achieving 10-15% better accuracy than GSQ-RCO on DeepSWE, Terminal Bench, Aider Polyglot and more, plus a new engram switching process supporting BF16/Q80/Q51/IQ4NL.
Key details:
- Rigorous benchmarking: 26.2B total tokens, with DeepSWE accounting for 18.1B; LiveCodeBench run 3x, HMMT and others 8-16x
- Settings: 256K context for agentic benchmarks, 128K otherwise; temperature 1.0, topp 0.95, topk 20
- RMSE vs BF16: Q80 at 0.0053 (54.4GB), Q51 at 0.0376 (38.4GB), IQ4NL at 0.0757 (28.8GB)
- Quants below UD-Q4KXL updated; larger ones planned
All quants are open on Hugging Face at unsloth/Qwen3.8-Flash-Next-GGUF.
Related event: Unsloth Releases Dynamic v3.1 Quantization for Qwen3.8-Flash-Next(4 posts)→
More from Infra
- Open-source TrnRT engine serves GLM-5.3-Flash on one Trainium2 box at 3.57M tokens/min — ycombinator · 2026-11-05
- Stanford study: NVIDIA B200 delivers 4x the dollar value per unit of energy vs AMD MI350X — rohanpaul_ai · 2026-11-05
- My Local AI opens Android closed test to run local models free on phones — TheOyinbooke · 2026-11-05
- Amazon CEO: Graviton and Trainium revenue commitments up more than 20x YoY — Beth_Kindig · 2026-11-05
- Semiconductor CEOs: ex-US demand more than sufficient if US buildout slows — inductionheads · 2026-11-05
- ZeroGR paper: filtering-first decoding speeds generative recommendation up to 9.9x — _reachsumit · 2026-11-05