I tried a new 8B local LLM, and its design might be the biggest shift since DeepSeek R1
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
Zyphra's Zaya1-8B introduces two novel architectural ideas: Compressed Convolutional Attention (CCA/CCGQA), which compresses queries, keys, and values into a shared latent space with convolutional mixing for a claimed 4x FLOP reduction at equal KV-cache budget, and Markovian RSA, which combines parallel reasoning traces with fixed-duration chunked reasoning to keep context bounded while enabling arbitrarily long reasoning chains. The model was pretrained on a 1,024-GPU AMD MI300X cluster with no Nvidia hardware. Running it locally proved difficult on AMD consumer GPUs due to kernel sizing mismatches (LDS overflow on RDNA3), but worked on Apple Silicon via an MXFP4 quant through vMLX at ~42 tokens/second. The full Markovian RSA scaffold is not yet available locally. On a hard math problem, the 8B model outperformed GPT-5.5 and Claude Opus, suggesting the architectural innovations meaningfully improve reasoning quality. Benchmark numbers are self-reported by Zyphra and the model is narrow, excelling at math and code over generalist tasks.
