August 30, 2026 ·

tencenthunyuanopen-weightscodingpricingbenchmarkschinaself-hostingstrategy

Tencent's Hy4 wins its own blind test 46.8% to 40.4%. That is the story.

TL;DR: On 28 August 2026 Tencent released and open-sourced Hy4 preview — 770B total parameters, 49B active, a 1M-token context, Apache 2.0 weights on Hugging Face, ModelScope, GitCode and CNB, and API pricing of $0.834 in / $2.501 out / $0.042 cached per million tokens. It is the fourth open-weight flagship in fifteen days and the first from a lab this site has never had cause to cover. The number that matters is not on the benchmark chart. Tencent ran a blind side-by-side evaluation — 163 of its own experts, 203 engineering tasks — and published the full breakdown: against GLM-5.3, Hy4 won 46.8%, tied 12.8%, lost 40.4%. Against Kimi K3, it lost 40.9%. A vendor’s own evaluators, on the vendor’s own tasks, judged a competitor better on two of every five. The open-weight frontier has converged to inside the noise. What still differs: price, licence, serving footprint, and what the model cannot do at all — Hy4 is text-only, and Tencent lists over-reasoning and over-verification as known defects. For you: the Apache 2.0 licence is real but 770GB of FP8 weights means most teams will rent this, not run it. Evaluate on price; do not plan around a 0.07-point capability margin.

What shipped

The specification is unusually well documented for a Chinese open-weight release, and all of it is on the model card rather than in a press summary.

Hy4 preview is a Mixture-of-Experts model: 770B total parameters, 49B activated per token, across 78 layers. The first layer uses a standard dense feed-forward network; the remaining 77 replace it with MoE, each carrying 256 routed experts and 1 shared expert, with every token activating the top-8 routed experts plus the shared one. A separate native MTP layer (10B total, 0.7B activated) is built in for speculative decoding. Context length is 1M tokens, hidden size 6144, 64 attention heads, vocabulary 120,832.

Weights are on Hugging Face, ModelScope, GitCode and CNB, in both full and FP8 variants, under Apache License 2.0. API access runs through Tencent Cloud TokenHub and OpenRouter at $0.834 per million input tokens, $2.501 output, $0.042 on cache hits. The model ships inside Tencent’s own WorkBuddy, CodeBuddy, Yuanbao and ima products, free on the first two for two weeks from launch.

It is text-only. There is no vision support, which for a 2026 flagship is a conspicuous omission and the single clearest limit on where it can be dropped in.

Most vendor launch material is a bar chart with the comparison set chosen to flatter. Tencent published something different, and it is worth quoting the structure exactly.

The company ran a blind side-by-side evaluation in which 163 internal experts — software engineers, game developers, finance analysts and security specialists — rated model outputs on 203 engineering tasks. The headline is that Hy4 came out ahead. The full result is this:

Comparison	Hy4 average	Rival average	Wins	Ties	Losses
vs GLM-5.3	2.99 / 4.00	2.92	46.8%	12.8%	40.4%
vs Kimi K3	2.99 / 4.00	2.94	51.2%	7.9%	40.9%

A 0.07-point margin on a four-point scale, and a competitor judged better on two of every five tasks by raters who work for the challenger. That is not a capability gap. That is a coin flip with a slight lean.

This is the most useful thing in the release, and Tencent deserves credit for not cropping it. It is also, read plainly, a lab telling its own market that the thing it just spent an enormous amount of compute on is interchangeable with what you already have.

The comparison Tencent did not run

The benchmark figures reported around the launch are stronger than the blind test, and the selection is instructive.

Tencent’s reported Terminal-Bench 2.1 score is 85.4, a 14.6-point gain over Hy3, framed as surpassing DeepSeek V4 Pro and matching Claude Opus 5. DeepSWE goes from Hy3’s 28.0 to 64.3. Other reported figures include GPQA 92.3, SWE-bench Pro 65.7, SWE Multilingual 82.9, Toolathlon-Verified 74.1 and APEX-Agents pass@1 37.1.

Now set those against the record. When Z.ai released GLM-5.3 on 14 August, it reported DeepSWE v1.1 at 66.9. Hy4 reports 64.3. Third-party aggregation places GLM-5.3 at 88.2 on Terminal-Bench 2.1 against Hy4’s 85.4, and independent measurement puts GPT-5.6 Sol at 88.8 on the same test.

So on the two benchmarks where a direct comparison is available, Hy4 appears to sit behind the model it narrowly beat in its own blind evaluation — and Tencent’s Terminal-Bench framing names Opus 5 and DeepSeek V4 Pro, but not GLM-5.3.

Two honest caveats, because they cut both ways. Benchmark versions and harnesses differ enough between labs that a two-to-three point cross-vendor gap is close to meaningless, so this is not evidence that GLM-5.3 is better. And that is precisely the argument: if the published numbers cannot separate these models, the published numbers are not a buying criterion. The same pattern collapsed under independent evaluation when Qwen 3.8 Max was measured by outsiders and landed tenth.

Convergence is architectural too

The model card credits its own lineage without hedging: the attention module uses Gated DeepSeek Sparse Attention with IndexCache, described as inspired by DeepSeek and GLM, with identity Hyper-Connections on the residual path.

That is a frontier lab stating that its flagship’s core efficiency mechanism comes from two competitors’ published work. Combined with Alibaba and Z.ai shipping near-identical cheap-tier architectures within hours of each other on 26 August, the picture is a tier where technique diffuses faster than any lab can build a lead on it.

Four open frontier releases in fifteen days — GLM-5.3 on the 14th, two Flash models on the 26th, Hy4 on the 28th — from labs borrowing each other’s attention mechanisms and landing inside a rounding error of each other. No individual open model is strategically important any more. The tier is.

The licence you probably cannot exercise

Apache 2.0 with no field-of-use or geographic restrictions is the most permissive licence in this class, and it is genuinely better than what Qwen’s restrictions turned out to allow. It is also, for most buyers, theoretical.

770 billion parameters at one byte each is roughly 770GB of FP8 weights before a single token of KV cache. Tencent’s official vLLM and SGLang recipes both launch at --tensor-parallel-size 8. An 8×H100 node holds 640GB and will not fit the model. You need H200-class cards (8×141GB = 1,128GB) or better, or multi-node tensor parallelism — and a 1M-token context wants real headroom on top.

Contrast that with the cheap tier: GLM-5.3-Flash at 320B total / 18B active, Qwen3.8-Flash-Next at 125B / 6B. Those are weights a mid-sized team can actually serve, which is what made self-hosting economics the real price floor rather than any published rate. Hy4’s open weights are, in practice, a licence to be audited and fine-tuned, not a licence to be independent — and that distinction matters more now that the distribution layer itself has an owner and US policy has been actively examining curbs on Chinese open-weight releases.

What actually separates the open flagships now

If capability is inside the noise, these are the axes that are not:

	Hy4 preview	GLM-5.3	Kimi K3
Licence	Apache 2.0	Open weights, held 2 weeks for safety hardening	Open weights
API price /1M	$0.834 / $2.501	~$1.40 / $4.40	~$3.00 / $15.00
Total / active params	770B / 49B	~743B	Large
Vision	None	Yes (5.3-Flash line)	—
Context	1M	—	—
Self-host realism	H200-class node minimum	Comparable	Comparable

Hy4 wins the price line clearly and loses the modality line clearly. That is a real decision, and it is a decision about fit, not about which model is smarter.

What to do this week

Evaluate on cost per merged change, not benchmarks. At $0.834/$2.501 a two-week trial through OpenRouter or TokenHub costs less than an engineer-day. Run it against real tickets from your own repository and measure completion rate and total spend — including the token overhead from the over-verification Tencent itself flags, which lands hardest on the long-horizon agentic work this model is sold for.
Check the modality gap before you plan a migration. Text-only rules Hy4 out of any pipeline that reads screenshots, diagrams or PDFs as images. If your coding agent takes visual input anywhere, this is not a drop-in.
Do not treat Apache 2.0 as vendor independence unless you have priced the hardware. Run the arithmetic on 770GB of FP8 weights against the GPUs you can actually get, in a quarter when AI server prices are rising on memory alone. If the answer is “we would rent it,” write that down — it changes your risk profile, not your capability.
Make open models fungible on purpose. Four converged flagships in fifteen days is a standing invitation to keep prompts, evaluations and tool definitions portable and to route through an abstraction rather than a vendor SDK — the same argument that made the neutral gateway layer worth something. Re-tender the slot quarterly. On current evidence the switching cost should be close to zero, and the pricing moves faster than that.
Keep the frontier comparison honest. Against Claude Opus 5 at $5/$25 or GPT-5.6 Sol’s promotional $4/$20, Hy4 is a sixth to an eighth of the price. That gap is worth a serious evaluation on cost-sensitive, high-volume, text-only workloads — and it is not evidence that the models are equivalent. Full context on the coding tier in the best AI coding tools guide and the DeepSeek and Qwen reviews.

Honest caveats

Every performance number here originates with Tencent. No independent evaluation of Hy4 preview has been published, and this site has run no tests of its own. The third-party figures cited for GLM-5.3 and GPT-5.6 Sol come from benchmark aggregators whose harnesses may not match Tencent’s.

It is labelled a preview, and the label is load-bearing. Tencent’s own words are that this is “an early version” with “real headroom left in both pre-training and post-training,” shipping with known issues — “spending longer than necessary reasoning through complex tasks, and a tendency to over-verify its own work.” Both directly inflate output-token spend, which is where the price advantage lives.

The blind evaluation was run by Tencent employees on Tencent-selected tasks. It is more informative than a bar chart because it publishes the losses, not because it is neutral. Treat the 40% loss rates as a floor on how close the field is, not as a measurement of anything else.

Update, 3 September 2026 — the convergence argument now has a counter-example running the other way. This article described a fifteen-day run of open flagship releases. The lab that built its reputation on exactly that has not joined it: Meta shipped Muse Spark 1.3 on 2 September as a proprietary model, and the weights for Muse Spark 1.2, pledged on 10 August to arrive “in the coming weeks,” remain unpublished. Meta’s only genuinely open release in the period is Muse Glimmer — 30B parameters under Apache 2.0, published 10 August, and not in the same capability class. The practical consequence for anyone assembling an open-weight shortlist is that the frontier of what you can actually download is still defined by the labs in this article plus Google’s Gemma, not by Meta. The exercisability point above applies with equal force in the other direction: a permissive licence you cannot run is limited, and a capable model whose weights never ship is not on the list at all. Why Meta’s open-weights promise keeps slipping.

Frequently asked questions

Should we switch our coding agent to Hy4 preview?

Evaluate it, do not switch on this news. The case for evaluating is genuinely strong: Apache 2.0 with no field-of-use or geographic restrictions, $0.834 input and $2.501 output per million tokens against $5/$25 for Claude Opus 5 and $4/$20 for GPT-5.6 Sol's promotional rate, and a 1M-token context. At those prices a two-week trial on your own regression suite costs less than a single engineer-day. The case against switching today is that every performance number attached to this model was produced by Tencent, and the comparison set was chosen. Tencent's Terminal-Bench 2.1 figure of 85.4 is presented against DeepSeek V4 Pro and Claude Opus 5 — not against GLM-5.3, which third-party aggregation places at 88.2 on the same test and which Tencent's own blind evaluation treats as its nearest rival. The model is also text-only and Tencent itself lists over-reasoning and over-verification as known defects, both of which inflate output tokens on exactly the long-horizon agentic work this model is pitched at. Measure cost per merged change on your real tickets, not benchmark deltas. If the price advantage survives the token overhead, it is large enough to justify the migration.

Is the Apache 2.0 licence actually useful to us?

Only if you can afford the hardware to exercise it, and most teams cannot. The licence is genuinely permissive — Apache 2.0 with no field-of-use carve-outs, unlike the restrictions that surfaced when independent evaluators examined Qwen 3.8 Max. But 770 billion parameters at one byte each is roughly 770GB of FP8 weights before you allocate a single token of KV cache, and Tencent's official vLLM and SGLang recipes both run at tensor-parallel-size 8. An 8×H100 node tops out at 640GB and will not hold the model. You need H200-class cards or better, or multi-node tensor parallelism, and then a 1M-token context wants substantial cache on top of that. Compare GLM-5.3-Flash at 320B total and 18B active, or Qwen3.8-Flash-Next at 125B and 6B — those are open weights a mid-sized team can genuinely serve. Hy4's are open weights you will almost certainly rent through Tencent Cloud TokenHub or OpenRouter. That is not a criticism of the licence; it is a warning against treating 'open' and 'runnable' as the same word when you write the vendor-independence section of your architecture doc.

What does the blind evaluation actually tell us?

More than the benchmark chart does, because Tencent published the losses. The setup was 163 internal Tencent experts rating outputs on 203 engineering tasks, blind and side by side. Against GLM-5.3 the result was 46.8% wins, 12.8% ties, 40.4% losses, averaging 2.99 to 2.92 on a four-point scale. Against Kimi K3 it was 51.2% wins, 7.9% ties, 40.9% losses, averaging 2.99 to 2.94. Read the win rate and Hy4 leads. Read the loss rate and a competitor's model was judged better on roughly two of every five tasks by evaluators employed by the company that built the challenger. A 0.05 to 0.07 point margin on a four-point scale across 203 tasks is not a capability gap you can plan a procurement around. The honest reading is that the top three open-weight models are substitutes for one another on engineering work, and that Tencent — unusually — gave you the data to see it rather than a bar chart with the losses cropped out.

How does Hy4 compare to Claude Opus 5 or GPT-5.6 Sol?

On price the gap is enormous and on capability the evidence is too thin to call. Hy4 lists at $0.834 input and $2.501 output per million tokens, roughly a sixth of Opus 5's $5/$25 and about an eighth of Sol's promotional $4/$20 on output. On capability, Tencent reports Terminal-Bench 2.1 at 85.4 and describes it as matching Opus 5; independent measurement puts GPT-5.6 Sol at 88.8 on the same benchmark, and third-party aggregation of SWE-bench Pro has Hy4 at 65.7 against figures for Sol that vary between 64.6 and 72.7 depending on the harness. That spread is the point: cross-vendor benchmark comparisons at this level of precision are not reliable, because harness, version and prompt scaffolding move scores by more than the gaps being claimed. What is reliable is that Hy4 is text-only while both frontier models accept images, that Hy4 ships as an explicit preview with acknowledged defects, and that nobody outside Tencent has yet reproduced its numbers. Treat it as a strong candidate for cost-sensitive, text-only, high-volume work and not as a drop-in replacement for a frontier flagship.

Why does it matter that a fourth lab shipped an open frontier model this month?

Because it changes what open weights are for. Through most of 2026 the open-weight tier was a small number of labs — DeepSeek, Moonshot, Z.ai, Alibaba — each with a recognisable position, and buyers tracked them individually. Fifteen days in August produced GLM-5.3 on the 14th, Qwen3.8-Flash-Next and GLM-5.3-Flash on the 26th, and now Hy4 from a lab with no prior presence in this coverage at all. Tencent's model card credits DeepSeek and GLM for the attention architecture it uses, which means the convergence is happening at the design level and not only in the scores. When four labs can independently reach the same capability band using each other's published techniques, no single open model is strategically important and the tier as a whole becomes a commodity input with a price near the cost of serving it. The practical consequence for a buyer is that the migration cost between open models should now be near zero by design. Keep your prompts, evaluations and tool definitions portable, route through an abstraction rather than a vendor SDK, and treat any individual open model as a line item you re-tender quarterly rather than a platform you commit to.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.