Tencent HY4 Preview Is Open-Source: 770B MoE, 1M Context, and a Price That Undercuts the Frontier

Tencent HY4 Preview, released and open-sourced on August 28, 2026, is the first genuinely competitive open-weight flagship from a Chinese lab in months, and its most surprising number is the price. Tencent lists it at 6 yuan (~US$0.85) per million input tokens and 18 yuan (~US$2.50) per million output tokens — roughly a tenth of what a top-tier closed frontier model charges for output — while publishing a 770-billion-parameter mixture-of-experts model, 49 billion of which are active per token, with a context window that clears a million tokens. It is the successor to the Hunyuan family's Hy3 (295B total, 256K context) and it quadruples that context while more than doubling active capacity. The weights are downloadable today from Hugging Face, GitHub, ModelScope, and GitCode, and the same model is already live inside Tencent's own products: WorkBuddy, CodeBuddy in both domestic and international builds, the Yuanbao assistant, and the ima workspace app. For anyone who has been watching open-weight models fall short of the frontier on software-engineering work, this is the release that finally makes the gap arguable — and the caveats are as important as the numbers.

The positioning is deliberate. Tencent frames Tencent HY4 Preview as a productivity model for four domains: software engineering, office and data analysis, game development, and scientific research. The engineering emphasis is what separates it from the generalist crowd — the release notes lead with long-horizon task understanding, planning, and debugging, not with chat quality. That focus shows up in the integration list as much as the benchmark table: CodeBuddy gets the model in its IDE, WorkBuddy gets it in office workflows (Tencent demonstrates it chewing through 72 financial documents in one pass), and game developers get MCP hooks into Unreal Engine 5 and Unity that turn a single sentence into a playable demo.

What actually shipped

The spec sheet is worth reading line by line, because each line changes what an open-weight model is allowed to do.

• Architecture — Mixture-of-Experts with 770B total parameters and 49B active per token, a roughly 16:1 sparsity ratio that keeps inference cost in a range a small team can actually afford.

• Context window — over 1 million tokens, four times Hy3's 256K. Long-horizon engineering tasks — a full repository in the window, a multi-file refactor, a large document batch — become single-request problems.

• Weights — MIT-style open release across Hugging Face, GitHub, ModelScope, and GitCode. This is a downloadable, self-hostable model, not a hosted teaser.

• API — available on Tencent Cloud's TokenHub, the vendor's own hosted endpoint, priced at 6 yuan per million input tokens, 18 yuan per million output tokens, and 0.3 yuan per million tokens for cache hits.

• Product integration — WorkBuddy, CodeBuddy (domestic and international), Yuanbao, and ima, meaning the same weights Tencent ships in its own apps are the ones you can pull and run.

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. What makes it more than a spec-sheet footnote is that Tencent says Tencent HY4 Preview found the bottlenecks itself — the model analyzed its own inference stack, and the optimizations are reported stable across context lengths and concurrency levels. 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 a preview that has been public for only a few days, it is also exactly the kind of claim nobody outside Tencent has verified 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 only a few days. Read them as the ceiling Tencent believes it hit, not as established fact.

The headline figure is Terminal Bench 2.1, a test of how well a model drives a real terminal while coding. Tencent reports Tencent HY4 Preview at 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, a software-engineering benchmark, jumps from 28.0 on Hy3 to 64.3, which Tencent describes as "gradually narrowing the gap" with first-tier models. On Toolathlon-Verified, a tool-calling test, it scores 74.1, which Tencent says surpasses Qwen 3.8 Max and GPT-5.6 Sol and approaches Kimi K3 and Claude Opus 5. On APEX-Agents pass@1 it lands at 37.1, a hair behind Kimi K3's 37.2. And 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 knowledge or reasoning, which is consistent with the productivity positioning rather than a coincidence. Second, Tencent is careful to describe DeepSWE and the others 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 the pricing actually buys

Pricing is where the release stops being interesting and starts being disruptive. At Tencent's list price, a million input tokens costs about what a single mid-tier API call costs on some platforms. The 18 yuan per million output tokens figure (~US$2.50) sits far below the $25-per-million output price of a top closed frontier model, and the 0.3 yuan cache-hit rate makes long agentic loops — where the same system prompt and conversation prefixes get re-sent constantly — unusually cheap to run.

This is also the one place where the routing layer changes the math. 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 own 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 like Tencent 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 — that is the mechanism that will carry Tencent HY4 Preview the moment it becomes routable, alongside the other open and closed models already in the catalog.

The parts that do not fit on a spec sheet

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

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 underlying operators that produced it. The 31.8% throughput gain is the most concrete public output of that loop: Tencent credits it to bottlenecks the model identified in its own serving stack. That is an initial recursive self-enhancement cycle: a model helping design the next version of itself. It is not unusual for a frontier lab to report pieces of this, but a Chinese open-weight model publicly describing the loop as a shipped capability is a step-change in how these releases talk about themselves.

The second is the mathematics result. Tencent reports that the model advanced the known volume 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's bound. Tencent also reports a 2.0x speedup on a 32,512-atom phospholipid bilayer simulation, down to 54.9 milliseconds per step, in the scientific-research lane. These are the kinds of results that usually get a model its own paper; here they are launch bullets, and like the rest of the launch material they are Tencent's own account.

What is missing, honestly

Tencent lists the known limitations plainly, and they matter for anyone deciding what to build on this model. 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 that the model shows 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, not even an Artificial Analysis entry. The model is too new for that. 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.

Who should pull the weights today

The honest answer is that this release separates cleanly into two audiences. If you run engineering workloads on your own infrastructure and the economics of closed APIs are what has been holding you back, Tencent HY4 Preview is worth a serious weekend test right now: the weights are open, the context window is repository-scale, and the list price makes a long agentic loop affordable in a way a $25-per-million-output model never will be. If you are running latency-sensitive production chat, or anything multimodal, or anything where you cannot afford a model whose only evidence is its vendor's own benchmarks, wait — the two-month iteration cadence Tencent has kept since February suggests a full Hy4 release is not far off, and the preview is explicitly an early iteration with known warts.

The practical middle path is a routing pattern: run the proven model on your critical path and the new one beside it, swapping in Tencent HY4 Preview on tasks where its cost profile wins and failing back automatically when it does not. That is the pattern a router exists for — the same OrcaRouter that carries Claude Opus 5, DeepSeek V4 Pro, and Gemini 3.1 Pro at their providers' list prices will carry this model the day Tencent opens it up. Until then, the weights are on GitHub, the API is on Tencent Cloud, and the claim that an open-weight model reached the top tier of agentic coding finally has a specific number attached to it. The number is Tencent's. The test is yours.