Unsloth rebuts GSQ-RCO: DeepSeek-V4.1-Flash 1bit AIME25 is 91.88%, not 100%
danielhanchen · x · 2026-10-30
Unsloth released new v3.1 GGUF quants for DeepSeek-V4.1-Flash-1017 with benchmarks showing they beat GSQ-RCO.
Key points:
- Could not reproduce GSQ-RCO's numbers: 1bit AIME25 measures 91.88%, not the claimed 100%; LiveCodeBench and GPQA claims are similarly off.
- Responding to claims that KLD can be gamed: minimizing KLD while minimizing disk size is the whole point of quantization; KLD of 0 means the quant equals the 16bit model.
- Benchmarks span AIME25, Aider Polyglot, LiveCodeBench v6, HMMT 2026, GPQA Diamond, MMLU-Pro, IFEval; lower-bit quants are shown to reason somewhat longer.
- Dynamic V3.1 leads nearly all benchmarks (see chart) and runs on 64GB+ RAM machines.
More from Models
- Polymarket puts Google at 59% to lead the AI race by year-end — Polymarket · 2026-10-30
- Nanbeige4.3-33B with 512K context tops agent benchmarks, real-world feedback sought — xdcfret1 · 2026-10-30
- LLMs struggle badly on tasks with more than one right answer — andpoul · 2026-10-30
- Nace AI open-sources Drex-Max v1: 27B multimodal decision model scoring 255 options in one pass — techNmak · 2026-10-30
- 0.87MB Hindi typing model Lekhani beats a 14x larger model in-browser — SuB8u · 2026-10-30
- Fake GPT-6.2 leak hoax hits 33k views in 6 hours, author reveals it was bait — Angaisb_ · 2026-10-30