Tencent Hy4 Ships 770B Open Weights Under Apache 2.0

A 770-Billion-Parameter Flagship, Released as Open Weights

On August 28, 2026, Tencent's Hunyuan team released Hy4 preview and put the weights on Hugging Face under an Apache 2.0 licence. It is the company's new flagship model, and it is also one of the largest open-weight releases of the year — 770 billion total parameters, of which only 49 billion activate on any given token.

That sparsity ratio is the headline for anyone who has to actually run the thing. A dense 770B model is a data-centre object. A mixture-of-experts model that lights up 49B parameters per token is a much more tractable serving problem, and the FP8 variant shipped alongside the main weights makes that explicit.

• 770B total parameters, 49B active per token — a mixture-of-experts design with 256 routed experts plus one shared expert per layer, top-8 routing
• Over 1 million tokens of context, using Gated DeepSeek Sparse Attention with cross-layer sparse index reuse
• Apache 2.0 licence, published on Hugging Face, ModelScope, GitCode and CNB
• API pricing of about $0.83 per million input tokens and $2.50 per million output tokens through Tencent Cloud TokenHub and OpenRouter

What Does Apache 2.0 Actually Change Here?

Licence terms are the least glamorous part of a model release and frequently the most consequential. Apache 2.0 is a permissive licence with no user-count thresholds, no acceptable-use rider bolted on, and no requirement to badge downstream products. For a model at this scale, that matters: it means a company can fine-tune Hy4 preview, serve it privately, and ship the result commercially without negotiating anything.

This is now a pattern rather than an exception. We have watched the same licence choice land on Kimi K3's 2.8T-parameter release and on Thinking Machines' 975B Inkling. The frontier of what you can legally download and run yourself keeps moving, and it is moving faster than the frontier of what you can afford to run.

How Well Does Hy4 Preview Actually Perform?

Tencent's own evaluation is worth reporting precisely because of how it was structured. The company ran a blind assessment with 163 domain experts across 203 engineering tasks. Hy4 preview averaged 2.99 out of 4.00, against 2.92 for GLM-5.3 and 2.94 for Kimi K3.

Two caveats belong right next to that number. First, this is a vendor-run internal evaluation, so treat it as Tencent's claim rather than an independent result until third-party benchmarks land. Second, Tencent's own published figures show the model still trailing GPT-5.6 Sol and Claude Opus 5 on the hardest reasoning suites, including Humanity's Last Exam and CritPt. The honest summary is that Hy4 preview sits at the top of the open-weight tier on practical engineering work, not that it has caught the closed frontier everywhere.

The more unusual detail is in the training write-up: Tencent says the model was used to help optimise parts of its own training and inference stack, and reports a 31.8% end-to-end throughput improvement against the baseline. Models tuning their own infrastructure is a quiet trend that deserves more attention than it gets.

Where It Fits Against the Rest of the Open Tier

The open-weight tier has become genuinely crowded in the past fortnight. GLM-5.3-Flash runs on 128GB of local RAM at a 3-bit quantisation, and Qwen3.8-Flash-Next fits in 75GB with no GPU. Hy4 preview is not competing in that bracket — at 770B parameters it is a server model first. What it competes on is being the largest permissively licensed option with a genuine million-token window, which is a different and useful thing to have available.

Why This Is Today's Biggest Story

A frontier-class model becoming downloadable changes what is possible for everyone downstream: research labs without frontier budgets, companies with data that cannot leave their own network, and the quantisation community that will spend the next fortnight making it fit on smaller hardware. That last group has a strong recent record — the same cycle played out within days for GLM-5.3 and Qwen3.8.

For readers tracking the broader picture, our AI coverage follows both sides of this: the models themselves, and the steadily improving toolchain that turns a 770B checkpoint into something you can actually put to work.

Sources: TechNode — August 28, 2026; Tencent Hunyuan on Hugging Face — August 28, 2026; LLM Stats release tracker — accessed September 1, 2026.