Tencent Hy4 Preview Is Open: A 770B MoE with a Million-Token Context at ¥6 per Million Input Tokens

Tencent Hy4 Preview, released and open-sourced on August 28, 2026, is the third Tencent flagship model in six months, and it is the one that stops being a research milestone and starts being a pricing event. It is a mixture-of-experts model with 770 billion parameters — 49 billion active per token — and a native context window of a million tokens, released under the Apache License 2.0 in both BF16 and FP8, downloadable today from Hugging Face, GitHub, ModelScope, and GitCode. And Tencent is pricing it at 6 yuan per million input tokens and 18 yuan per million output tokens, around US$0.83 and US$2.50, which puts a top-tier open-weight model below nearly every closed frontier model on the market. One release, three faces: open weights you can run yourself, an API on Tencent Cloud's TokenHub, and the same weights already live inside Tencent's own products, WorkBuddy, CodeBuddy, Yuanbao, and ima. Here is what actually shipped, what Tencent claims it scores, and what nobody has verified yet.

The engineering story matters more than the parameter count, because each piece is a lever on cost. The model is 78 layers deep: the first layer is a standard dense feed-forward network, and the remaining 77 are MoE layers with 256 routed experts plus one shared expert each, activating the top-8 routed experts per token. That roughly 16:1 sparsity ratio is what makes a 770B model affordable to serve — only a 49B slice does work on any given token. A dedicated multi-token-prediction layer (10B parameters, 0.7B active) enables speculative decoding, and the attention stack uses gated DeepSeek Sparse Attention with an IndexCache to keep long-context memory use in check. Reasoning is switchable between a long-thinking code>high/code> mode and a token-saving code>no_think/code> mode. These are the details that determine whether the headline price survives contact with a real workload, not whether the model can write a haiku.

The spec sheet, annotated

Read the release notes line by line, because each line quietly changes what an open-weight model is allowed to do.

• Parameters — 770B total, 49B active per token, a ~16:1 sparsity ratio that keeps serving cost in the range a small team can actually afford.

• Context — a native 1,048,576-token window, four times Hy3's 256K. A full repository, a multi-file refactor, a 72-document financial package — single-request problems.

• License — Apache License 2.0 for both the BF16 original and the FP8 quantized release, which means self-hosting, fine-tuning, and commercial use without a proprietary clause.

• Inference stack — Tencent lists vLLM and SGLang as the supported serving frameworks, the two the open-weight ecosystem actually standardizes on.

• API — Tencent Cloud's TokenHub endpoint at ¥6 per million input tokens, ¥18 per million output tokens, ¥0.3 per million on cache hits.

• Product integration — the same weights Tencent ships in WorkBuddy, CodeBuddy (domestic and international), Yuanbao, and ima are the ones you can pull and run. CodeBuddy gets a two-week free trial of the model on launch.

Tencent also reports an inference-engineering figure that rarely makes a launch note: a 31.8% end-to-end throughput improvement from operator fusion and communication optimization. That is the kind of detail that separates a model a lab ships as a research artifact from a model a company expects production traffic to hit. On day one, it is also exactly the kind of claim nobody outside Tencent has measured yet.

The scores worth quoting

Every benchmark Tencent published for Tencent Hy4 Preview is vendor-reported — run on Tencent's own evaluation setup, unreproduced by any independent lab, and the model has been public for less than a day. Read them as the ceiling Tencent believes it hit.

The headline figure is Terminal Bench 2.1, a test of how well a model drives a real terminal while coding. Tencent reports 85.4, which it says ties Claude Opus 5 and passes DeepSeek V4 Pro, and which beats its own predecessor Hy3 by 14.6 points. That single number is doing a lot of work — it is the claim that puts an open-weight model at parity with the most expensive closed frontier models on a hard agentic-coding test — and it is the claim most in need of independent confirmation. The rest of the internal table: DeepSWE jumps from 28.0 on Hy3 to 64.3, which Tencent describes as "gradually narrowing the gap" with first-tier models; Toolathlon-Verified, a tool-calling test, lands at 74.1; APEX-Agents pass@1 at 37.1, a hair behind Kimi K3's 37.2; SWE-bench Pro at 65.7 and SWE-bench Multilingual at 82.9. In a separate internal blind test, 163 experts scored 203 engineering tasks at 2.99 out of 4.00, edging GLM-5.3's 2.92 and Kimi K3's 2.94.

Two patterns stand out. First, every strong number is on agentic engineering work — terminal driving, tool calling, repository-scale tasks — and none is on generic chat or knowledge, consistent with Tencent's positioning of the model as "built for productivity" across software engineering, office and data analysis, game development, and scientific research. Second, Tencent is careful to describe DeepSWE and the rest as narrowing, not closing, gaps. The one clean, marketable parity claim is the Terminal Bench 2.1 tie, and it is the claim most worth testing with your own workload.

What ¥6 per million actually buys

Pricing is where the release stops being interesting and starts being disruptive. At 18 yuan per million output tokens — about US$2.50 — a long agentic coding session that burns a million output tokens costs roughly a tenth of what the same session costs on a top closed frontier model, and the 0.3-yuan cache-hit rate makes the re-sent system prompts and conversation prefixes of an agentic loop dramatically cheaper than most competitors' caching. Tencent's own demo has the model chewing through 72 financial documents in a single pass; at these rates that kind of batch is an affordable daily job instead of a budget line.

This is also where the routing layer changes the math, and it is worth being precise about it. OrcaRouter does not route Tencent Hy4 Preview yet — Tencent has not opened this model to third-party inference platforms, so it runs on Tencent Cloud's TokenHub endpoint and on self-hosted deployments of the open weights, and we do not claim to serve what we do not. But the discipline that makes a price cut matter is the same one the rest of the catalog runs on: OrcaRouter passes provider list prices through with zero markup, so when a vendor prices a token at 6 yuan, the number you pay on our side is 6 yuan, not a resold and marked-up version of it — and the day a vendor changes a price, the change is live on our side the same day. One API for 200+ models, automatic failover across providers when one hiccups, and a routing DSL to compose models into a single call: that is the mechanism that will carry Tencent Hy4 Preview the moment Tencent opens it up, and it is the mechanism you can already use to run a new, unproven model beside a proven one without betting your production path on either.

The claims that deserve a second read

Two things in the launch material are about how the model was built, not what it scored, and they deserve separate attention.

The first is the self-improvement loop. Tencent says Tencent Hy4 Preview participated in its own research and development — it was used to optimize the training methods, data strategy, evaluation systems, and low-level operators that produced it. That is an initial recursive self-enhancement cycle: a model helping design the next version of itself, and Tencent publishing it as a shipped capability. It is a big claim, and it is entirely self-reported.

The second is a mathematics result. Tencent reports the model advanced the known lower bound of the Blaschke–Lebesgue problem — a long-standing open question in convex geometry about the minimum-volume body of constant width — from 0.380799 to 0.41104, bringing it within roughly 2% of the Meissner tetrahedron conjecture. Tencent also reports a 2.0x speedup on a 32,512-atom phospholipid bilayer simulation, down to 54.9 milliseconds per step. Results like these usually get a model its own paper; here they are launch bullets, and like everything else on day one they are Tencent's own account.

The honest gaps

Tencent lists the known limitations plainly. There is no vision or multimodal input — Tencent Hy4 Preview is a text model, full stop, so anything image- or video-adjacent belongs on a different model. Tencent also flags slow-start behavior on complex tasks — long thinking before first output — and a tendency to over-self-verify, burning tokens double-checking work it has already done. That combination is exactly wrong for latency-sensitive chat and exactly tolerable for offline batch engineering work, which tells you where Tencent expects you to run it.

And the most important absence is verification. There are no independent scores for Tencent Hy4 Preview anywhere yet — not on a public leaderboard, not from a third-party eval lab. The model is too new. The internal expert blind test is a useful signal about how Tencent's own reviewers see it against GLM-5.3 and Kimi K3, but it is Tencent's reviewers on Tencent's tasks. Everything above the spec sheet is a claim awaiting a check.

How to actually try it today

The practical path is short. Pull the weights from Hugging Face, GitHub, ModelScope, or GitCode and serve them with vLLM or SGLang; call Tencent Cloud's TokenHub API if you want a hosted endpoint at the list price; or use it inside CodeBuddy or WorkBuddy, where Tencent is running a two-week free trial. The Hy3 model's free access runs through September 30, and Tencent's own cadence — roughly one major iteration every two months — points to a full Hy4 release not far behind the preview.

The honest summary: an open-weight, 770B, million-context model at ¥6 per million input tokens is a real thing that exists as of today, and the strongest benchmark attached to it is a vendor's claim about a tie with a closed frontier model that nobody has reproduced yet. Pull it, run the test you actually care about, and let the evidence decide whether the price stays the story.