On August 28, Tencent released and open-sourced Hunyuan Hy4 preview, with 770 billion total parameters and 49 billion activated per token, and a native context window of 1,048,576 tokens. The model weights are available under the Apache 2.0 license on Hugging Face, ModelScope, and GitCode, and the model is also live on Tencent Cloud’s TokenHub, OpenRouter, and Tencent’s own WorkBuddy/CodeBuddy, Yuanbao, and ima. Tencent describes it as “an early version of the Hy4 iteration,” with the official line being that it’s “firmly in the first tier of open-source models.”
According to the model card, Hy4 preview is a classic high-sparsity mixture-of-experts model: of its 78 layers, only 1 is a dense feed-forward layer, while the remaining 77 route through experts — each layer has 256 routed experts plus 1 shared expert, and each token selects 8 routed experts. The hidden layer width is 6,144 and the vocabulary size is 120,832. The attention mechanism uses gated sparse attention with IndexCache, compressing queries to 2,048 dimensions and key-values to 512 dimensions, paired with 4 iHC residual streams. On the deployment side, it supports vLLM and SGLang, with MTP speculative decoding enabled by default.
Behind the 6.4% Activation Ratio Is a Compute Ledger
49 billion divided by 770 billion works out to an activation ratio of roughly 6.4%. That number is the key to understanding the model’s pricing: total parameter count determines memory and loading cost, while the activated parameter count determines the actual compute per token. Tencent has pulled the two more than 15x apart, and the payoff is a price list that isn’t especially expensive: 6 yuan (about $0.834) per million input tokens, 18 yuan (about $2.501) per million output tokens, and 0.3 yuan for cache hits.
That pricing sits in the mid-range of China’s open-source camp. GLM-5.3-Flash, released the same week, prices input as low as $0.075 per million tokens — one-eleventh of Hy4 preview’s rate — but that’s an 18-billion-activated lightweight tier. Hy4 preview is aimed at heavier work: reading code across files, analyzing across documents. The two aren’t competing in the same lane, so comparing prices head-to-head doesn’t mean much.
The real cost variable is hidden in the 1M-token context window. Filling a million tokens of input once costs roughly 6 yuan at list price, dropping to 0.3 yuan on a cache hit — the cached price is just one-twentieth of the original. That size of discount is basically Tencent telling developers: treat long context as a standing workspace, don’t retransmit it every time.
How to Read That 0.07-Point Gap
Tencent’s blind-test results: 163 internal experts scored 203 engineering tasks, with Hy4 preview averaging 2.99 out of a possible 4, versus 2.92 for GLM-5.3 and 2.94 for Kimi K3.
The score gap is small, but the win/loss breakdown is more telling. Against GLM-5.3, Hy4 preview won 46.8% of matchups, tied 12.8%, and lost 40.4%; against Kimi K3, it won 51.2%, tied 7.9%, and lost 40.9%. In both pairings, the loss rate sits around 40% — meaning the opponent did better on nearly half of the individual tasks. All three models are bunched tightly together, with none pulling ahead.
Two caveats shouldn’t be skipped: the judges were Tencent’s own internal experts and the task set was also selected by Tencent, so this kind of self-organized blind evaluation carries an inherent home-field advantage; and with a sample of 203 tasks, a gap of 0.05 to 0.07 points carries fairly limited statistical significance.
Public benchmark numbers offer a more comparable picture: 82.9 on SWE-bench Multilingual, 65.7 on SWE-bench Pro, 85.4 on Terminal-Bench 2.1, 92.3 on GPQA Diamond, and 55.4 on HLE with tools. SWE-bench Pro is consistently where models take their biggest hit, and 65.7 puts it in the top tier among current open-source models.
Betting on Long-Horizon Work
Tencent is explicit about where it’s aiming this model: long-horizon software engineering, cross-file office analysis, game development, and scientific research. The model offers two reasoning modes, high and no_think — the former for scenarios that need extended chains of reasoning, the latter to save on tokens.
One example Tencent points to as a showcase is the 3D Blaschke–Lebesgue problem, where the model pushed the volume lower bound from 0.380799 to 0.41104 — still about 2% short of the Meissner tetrahedron conjecture’s 0.41986. That kind of example is only so persuasive, since it’s a hand-picked showcase, but the direction is clear: Tencent wants Hy4 preview working inside research pipelines, not just sitting in a chat window.
For developers in China, the practical upshot is one more option on the table. GLM-5.3, Kimi K3, and DeepSeek-V4-Pro have already crowded the open-source long-context space, and Hy4 preview is squeezing in on the strength of its permissive Apache 2.0 license plus Tencent Cloud’s ready-made inference service. The license matters in particular — many domestic open-source models come with commercial-use restrictions attached, whereas Apache 2.0 means enterprises can modify, deploy, and resell it directly without negotiating a separate license.
A preview is still a preview. Tencent hasn’t given a timeline for a formal Hy4 release, nor said how different the preview and the eventual release will be.
Sources: Tencent Hunyuan technical blog, Hugging Face model card, IT Home, CocoLoop, DataLearner. Model parameters, blind-test methodology, and API pricing verified against the official model card and pricing page; price conversions are rough estimates at list price.