Hy4 preview arrives as a practical step toward production-ready LLMs for long-horizon engineering, multi-file office workflows, game prototyping, and STEM research. Rather than being an academic showcase, its design choices prioritize maintaining usable inference at million-token scale while giving developers access to deployable BF16/FP8 checkpoints and deployment recipes.
Key Capabilities
- Large-scale MoE backbone with practical activation: the model is a 770B-parameter Mixture-of-Experts design that activates ~49B parameters per token (top-8 routed experts + shared expert), enabling higher capacity without linearly increasing per-token compute cost — so you get stronger reasoning and coding capability for long contexts.
- Native million-token context and sparse attention: Gated DeepSeek Sparse Attention plus cross-layer IndexCache support 1,048,576-token windows, which means workflows spanning many files or long agent traces can be processed in a single pass instead of being repeatedly truncated.
- Speculative MTP decoding and deployment-first tooling: an integrated MTP speculative-decoding layer plus published vLLM/SGLang recipes and FP8/BF16 weights reduce the engineering lift to serve the model in production-like settings.
- Open-source, deployment-ready weights: Apache-2.0 licensed checkpoints (BF16 and FP8 quantized) and recommended inference parameters (e.g., temperature=0.9, reasoning modes) let teams evaluate and integrate the model without restrictive licenses.
Who it's for and tradeoffs
Great fit if you need a research- or product-focused LLM that must handle extremely long contexts (multi-file engineering, document analysis, long agent traces) and you have cluster resources to host large MoE checkpoints. It’s especially relevant for engineering teams wanting lower-latency iterative workflows via speculative decoding and engineers who can integrate vLLM/SGLang.
Look elsewhere if you need a lightweight on-device model, minimal infra cost, or a fully stabilized production-released model; Hy4 preview is explicitly an early preview with known tendencies to over-verify and spend extra time on complex multi-step reasoning, and practical deployment typically requires multi-GPU server setups and engineering effort.
