Sergii Kozyrev and Davyd Maiboroda release Minima
Minima released — to test the practice, in early community 4-bit quantizations of the hybrid LLM Qwen3.8-27B (48 GDN layers, 16 attention layers), of leaving the GDN blocks — especially the decay and write-strength gates — at 8/16 bits for fear of accumulating recurrent error, applies NVFP4 W4A4 to all 496 linear layers. Reports that it matches BF16 within the range of seed noise (an average of -0.52 across 5 tasks) on 4K/32K perplexity, MMLU-Pro, GSM8K, AIME'25, GPQA-Diamond, LiveCodeBench and 64K RULER retrieval, while at 17.5 GiB it is the smallest of those compared and its prefill is 14–19% faster. Also fixes the global-scale mismatch that arises when a kernel that fuses modules into a single GEMM serves NVFP4 checkpoints calibrated per module
Why it matters
Overturns the conventional wisdom that "errors in recurrent state accumulate over long contexts", arguing that the recurrent half of a hybrid LLM is in fact the easier half to quantize, and offers four mechanistic explanations, including that NVFP4's 16-element block scaling localizes extreme outliers in the residual stream and that the delta-rule recurrence keeps injected noise flat across 32K tokens. Figures are as reported by the authors.
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/2609.04098Added in Week 36, 2026 · See that week