Unsloth burns ~30B tokens to show IQ_3S quant of Qwen3.8-Flash-Next hits 99% of BF16
DigiDecode_ · reddit · 2026-11-05
Unsloth released UD v3.1 dynamic GGUF quants for Qwen3.8-Flash-Next, claiming 15% better accuracy than GSQ RCO at similar size. The bigger takeaway for local users: IQ3S GSQ RCO scores 99.0% of the BF16 baseline averaged over 12 benchmarks (DeepSWE v1.1, Terminal-Bench 2.1, LiveCodeBench v6, etc.), with Unsloth spending 30B tokens on evaluation. UD-IQ3XXS recovers 99.9%. Existing IQ3S+Strata setups are fine as-is; new downloads should grab UD v3.1.
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