Unsloth Dynamic v3.1 GGUFs cut tokens ~50% and beat other quants on Qwen3.8 and DeepSeek-V4.1
danielhanchen · x · 2026-11-12
Unsloth released Dynamic v3.1 GGUFs for Qwen3.8-Flash-Next and DeepSeek-V4.1-Flash, targeting quantized models' overthinking:
- Qwen3.8-Flash-Next: thinking and token use brought to 95% of full precision; up to 50% fewer tokens, 7% faster, and 15% higher accuracy vs other same-size quants, benchmarked at 512K context over 30B tokens and 12 tasks incl. Terminal Bench and DeepSWE
- DeepSeek-V4.1-Flash-1017: clearly beats GSQ-RCO quants on agentic benchmarks (DeepSWE, Terminal Bench 2.1, Automation Bench, Aider Polyglot) at 512K context, with many tiers near or above the 8-bit baseline
- Hardware needs (DeepSeek): 1-bit 128–145GB, 4-bit 305GB, 8-bit 315GB
A useful reference for local deployment and token-efficient inference.
More from Infra
- Dell expands AI Data Platform: agents have a data-context problem, not a model problem — DavidLinthicum · 2026-11-12
- Potion UI's local login system turns your PC into a multi-user AI server — cocktailpeanut · 2026-11-12
- Recursive's Eureka Machine ships Flint: a sparse 40B model with 2.5B active for local agentic coding — agihouse_org · 2026-11-12
- AWS Finally Built Its Own AWS Emulator — super_linear · 2026-11-12
- Megakernel is multiprogramming for GPUs: 'GPUs are entering the CPU age' — pranjalssh · 2026-11-12
- Gemini Decision 1 costs $345 per 1M tasks, 3x pricier than leader Drex Max v1.1 — ArtificialAnlys · 2026-11-12