Most small-parameter LLMs trade reasoning ability for footprint; ZAYA1-8B flips that expectation by using a Mixture-of-Experts design plus a post-training reasoning phase so a model with relatively few active parameters can match much larger models on formal math and coding benchmarks. The release targets scenarios where test-time compute and latency matter but you still need strong stepwise reasoning.
AI Model2026
ZAYA1-8B
Mixture-of-Experts LLM tuned for mathematical and coding reasoning, with ~760M active / 8.4B total parameters and post-training for improved stepwise reasoning. Optimized for inference efficiency (vLLM/transformers forks) so it can run in computation-constrained or local deployments; Apache-2.0 licensed.
Introduction
Information
- Websitehuggingface.co
- AuthorsZyphra
- Published date2026/05/04
Categories
More Items
AI Model2026
Answers structured decision questions (choice / yes-no / score) about text, images and videos in one call and returns probabilities for every option. A 27B multimodal model (includes a 0.46B vision encoder), supports zero-shot classification, image/video inputs and integrates with vLLM / vLLM Semantic Router for high-throughput serving. Apache-2.0.
AI Model2026
Generates calibrated decisions (choice/score/noul) from a prompt and optional images, up to 256 options. Routed MoE experts are quantized to NVIDIA NVFP4 to reduce weights to ~25.6 GB and GPU memory to ~23.3 GiB; provided as a vLLM-compatible System 1 decision model with a /v1/decide endpoint.
AI Model2026
Runs a 24B-parameter mixture-of-experts reasoning model with only ~2B active parameters per token for fast on-device agentic workflows; provides a 131K-token context and 128K vocabulary, post-trained for tool calling and agentic RL—pair with retrieval for knowledge-heavy tasks.