Paper: https://arxiv.org/abs/2611.08376 | HF: https://huggingface.co/papers/2611.08376

At 10M tokens, a 1.3B Transformer needs 2.1 TB of KV cache (0.32 TB even at 2.5-bit quantization). Channel-wise gated delta models (like Gated DeltaNet-2 / GDN-2) compress history into a fixed 25 MB state, but hit three bottlenecks across 100B-token training that Muir resolves:

• 100x slow-channel glare -> Matched-filter readout: Slow-decaying rows accumulate up to 100x more energy than fast rows, swamping query readout and forcing harsh write-time erasure that overwrites early facts. Equalizing rows via an inverse fractional power law + per-token query modulation relaxes write erasure by 17% and extends lookback horizon by 60%.
• 13x write-vs-erase sensitivity -> +30.4% wider FFN: Normalized keys concentrate onto ~5% of state rows while values stay diffuse, making updates 13x more sensitive to value writes than key erasure. Fusing all 5 gates into one rank-512 projection cuts gate parameters 4.5x to widen the FFN by +30.4% at strict parameter budget.
• Numerically-stable, large-chunk parallel kernel: Unpivoted decay division underflows fp32 and NaNs 25% of heads at large chunk sizes. Boundary-pivoted multiplicative decays stay in (0, 1] with zero division, while a recursive block-Neumann solver routes 98% of chunkwise FLOPs onto matrix units (24x fewer vector FLOPs).​

https://preview.redd.it/zgsr9oynm10i1.png?width=2382&format=png&auto=webp&s=9373d5a667567c211f679bdf0759ed8dcc79dbd3

https://preview.redd.it/unnegoynm10i1.png?width=2289&format=png&auto=webp&s=c6390b2a2324527fcb15576e1614bc4a3349a961

Results:

• Quality and speed: Beats GDN-2 and Transformer on perplexity and zero-shot reasoning at every scale (1.2x token efficiency vs. GDN-2), and trains 2.7x faster than Splash Attention at 32K.
• Recall beyond passkey: Compared to GDN-2, lifts RULER 5-hop Variable Tracking root retention (28.4% -> 48.0%), early-depth recall at 8K (1% -> 31%), and real-world extractive QA (+2.5 avg pts).
• 10M streaming and 1M passkey: Streams 10M tokens without perplexity drift in a 25 MB state (dynamic readout), and hits 90.0% zero-shot / 99.4% post-cooldown passkey accuracy at 1M tokens (static readout, where GDN-2 collapses past 4K-8K and Transformer OOMs past 64K).