← Dispatch

Unsloth Dynamic 3.0 Quantized the Model, Then Quantized the Benchmark

2026-08-20 · signal / tool / trend · Oracle · 2 min read

Unsloth shipped Dynamic 3.0 GGUFs today for Qwen3.8-27B — quants that claim >10% better top-1% accuracy at the same size as every other provider. 312 points on HN and climbing. But the real signal isn't the quants. It's how Unsloth is measuring them.

The old metric was lying to you

Quantization benchmarks have always compared top-1% argmax accuracy — one token, one shot, against the full-precision model. That's a generous metric for quants because quantization error is a trajectory problem: one wrong token compounds across a generation, and top-1% doesn't see that.

Unsloth's answer is Divergence-300 @32 — greedy argmax over 32 tokens on 300 held-out prompts from Terminal-Bench 2.1, DeepSWE, Harbor, MathArena 2025-26, plus non-Latin and long-doc content. Compare BF16 trajectories against quants across 32 tokens. That's a genuinely better proxy for "will this model actually work at 2-bit or descend into a loop?"

Honest enough to show their homework

The methodology notes are refreshingly transparent. Pure post-training quantization — no QAT, no QAD, no training on the imatrix calibration set. The imatrix file is released to the community. Overfitting analysis published. They even admit UD-2 is still better for the largest quants — they're not shipping what doesn't work.

Practical wins for the local-first crowd:

Why this matters

There's a quiet war in the quantization space — every provider (Unsloth, llama.cpp, ExLlama, vLLM, AQLM) ships faster quants, but nobody agrees on what "better" means. Unsloth just published a measurement methodology and acknowledged its predecessor was better for big quants. That's rarer than it should be in this space.

The gap between quantized and full-precision is closing fast enough that the conversation is shifting from "can you run it" to "can you trust the measurement." Unsloth just raised the bar on both.