Tencent drops Hy3, a 295B open-source MoE built for agent coding

Tencent just open-sourced Hy3, a 295B-parameter Mixture-of-Experts (MoE) model designed squarely for reasoning and agentic programming tasks. While its total footprint sits at 295B with a 3.8B MTP layer, it only activates 21B parameters per token, packing a massive 256K context window under an Apache 2.0 license.

Alongside the base weights, Tencent also released a ready-to-run FP8 quantized version across Hugging Face, ModelScope, GitCode, and CNB.

What it does

Hy3 is tailored for real-world developer workflows rather than just chasing generic benchmarks. It handles full-stack frontend design, CI/CD pipeline automation, complex financial modeling, and vibe-coding full web applications or games.

In an internal blind evaluation involving 270 domain experts running actual workplace tasks, Hy3 scored 2.67 out of 4, edging past GLM-5.1 at 2.51. Its lead was most noticeable in frontend development, data storage, and CI/CD pipelines.

On hard programming evals, Hy3 posted a 78 on SWE-bench Verified, 57.9 on SWE-bench Pro, and 90.4 on GPQA Diamond. Crucially for tool builders, it stays consistent across different agent frameworks: accuracy variance across scaffoldings like CodeBuddy, Cline, and KiloCode stays within a tight 4% margin.

To address common agent failure points, Tencent tuned post-training data and RL constraints aggressively. In internal real-world testing, Hy3 cut hallucination rates from 12.5% down to 5.4%, commonsense errors dropped from 25.4% to 12.7%, and multi-turn intent tracking failure rates fell from 17.4% to 7.9%.

How to run it

Because Hy3 carries 295B total parameters, running it requires serious metal. Even the FP8 quantized build takes up around 300GB of disk and memory storage. For production serving, Tencent recommends an 8-GPU node using high-memory cards like H20-3e.

Serving is streamlined through official recipes in both vLLM and SGLang, each supporting multi-token prediction (MTP) acceleration out of the box. Once launched, you interact with it through a standard OpenAI-compatible API.

You can also dynamically toggle reasoning effort: pass no_think for immediate chat responses, or switch to high when asking the model to tackle complex math, debugging, or deep logical chains.

Full model weights, FP8 checkpoints, fine-tuning scripts, and verl-based RL recipes are available now on Hugging Face.

Why it matters

Hy3 shows that a 295B MoE activating just 21B tokens can go toe-to-toe with trillion-parameter giants on software engineering tasks without demanding an impossible runtime budget.

Tencent's decision to optimize for scaffolding invariance and low hallucination rather than synthetic leaderboards is refreshing for developers building production coding agents. It is too heavy for casual hobbyist laptops, but for engineering teams looking to host their own private, high-accuracy coding assistant, Hy3 is one of the strongest open-weight foundations available right now.