Unsloth says it can't reproduce GSQ-RCO results, ships DeepSeek-V4.1 Dynamic v3.1 GGUFs
danielhanchen · reddit · 2026-10-30
Unsloth released preview Dynamic v3.1 GGUFs for DeepSeek-V4.1-Flash-1017 and benchmarked them against ISTA's GSQ-RCO quants across 7 benchmarks, saying it could not reproduce GSQ-RCO's reported results: the claimed 100% AIME25 accuracy recovery measured only 91.88%. UD v3.1 beats GSQ-RCO on nearly all metrics (AIME25, LiveCodeBench, GPQA Diamond, MMLU-Pro, Aider Polyglot), and even the older v3 beats ISTA; each benchmark ran 3+ times (AIME25 16 times). Low-bit quants also reason somewhat longer, but UD leads there too. Runs on 128GB RAM devices, or 64GB with Engram offloading off. Unsloth Desktop also gains Browser Use, auto compaction, a TTS pipeline with voice cloning, and sandboxing. Full data at unsloth.ai; weights on Hugging Face.
More from Infra
- NVIDIA's Nemotron 3.5 Super hits 51.78% on SWE-bench Pro with 4.6x fewer active params — dair_ai · 2026-10-30
- Bonsai Compresses Qwen 27B from 55.6GB to 5.9GB, Retains 98.2% Benchmark Performance — nicolechirps · 2026-10-30
- Weaviate: Use Slow-Query Log and Profiling to Fix Slow Hybrid Search Queries — CShorten30 · 2026-10-30
- Nearly $1 Trillion Wiped From US Stocks at Open as AI Chip Stocks Lead Selloff — CurieuxExplorer · 2026-10-30
- Anthropic eyes $2T valuation on $100B raise, with $431B compute commitments and circular vendor financing — LexSokolin · 2026-10-30
- Cloudflare private origins goes GA: WAF and caching for internal APIs and MCP servers — Cloudflare Blog · 2026-10-30