Hy4 Preview is a 770B MoE model (49B active, 1M context) from Tencent, built for long-horizon agentic tasks. It autonomously handles coding, game dev, and complex document analysis, running its own tests and fixing bugs before delivery.
Maker
indie_inkwell
Loading comments...
Hy4 preview is a large-scale Mixture-of-Experts (MoE) language model from Tencent, positioned as a foundation for long-horizon agentic work. It carries 770B total parameters with 49B active at inference, paired with a 1M-token context window — a combination aimed squarely at tasks that run long, branch often, and demand sustained coherence.
Rather than behaving like a passive chat assistant, Hy4 preview is designed to act. It autonomously handles coding, game development, and complex document analysis, running its own tests and fixing bugs before delivery. That self-verification loop is the core idea: the model doesn't just produce output, it checks the output and iterates until the work holds up.
Hy4 preview uses a Mixture-of-Experts design totaling 770B parameters, with 49B active during inference. This structure aims to deliver high capability while keeping the active compute footprint more manageable than a dense model of equivalent scale.
The model supports a context length of up to 1 million tokens. That capacity is what makes sustained, long-horizon agentic tasks feasible without constantly losing earlier context.
Hy4 preview is built to take on coding and game dev tasks end to end. It runs its own tests and fixes bugs before delivery, closing the loop between generation and verification.
Beyond code, the model is designed for in-depth analysis of complex documents. The long context and agentic behavior combine to support review work that spans substantial material.
Hy4 preview doesn't just generate — it tests its own work and fixes what breaks before handing anything over.
That self-correcting behavior is the differentiator. Most models stop at producing a plausible answer; Hy4 preview is framed around delivering verified results, which matters enormously when the output is code or a multi-step agentic plan. The 1M context and MoE scale exist in service of that goal — they give the model room to work and the capacity to do it well.
You're building agents that need to run for a long time without losing the thread, or you want a model that treats testing and bug-fixing as part of the job rather than an afterthought. If your work involves large codebases, game projects, or dense documents that won't fit in a typical context window, Hy4 preview is aimed directly at that space.
Other tools you might consider
Meet Mellum, a family of fast language models, including a next-generation model for ultra-low-latency and high-performance inference.
Clears is an agentic execution platform for autonomous software delivery, helping R&D organizations move beyond individual AI tools to agentic execution across the entire SDLC.
GitHits gives coding agents access to the open-source code your app depends on. Get real implementation examples, dependency source navigation, package inspection and documentation. Agents can grep and read your codebase. They can't grep and read the open-source code your app depends on. That's where they start guessing, retrying, and looping. GitHits builds a version-aware index on demand. Agents can search, navigate, and inspect the code behind their dependencies. CLI: npx githits@latest init
Harness Starter Kit helps teams turn fragile AI coding prompts into durable repository rules: AGENTS.md, drift checks, failure memory, adoption reports, and stack profiles for safer agent collaboration.
Loading comments...
