Tencent’s Hy4 Is the Fourth-Best Open-Weight Model on Paper, but the Weights Are Still Missing
Tencent's Hy4 ranks fourth among open-weight models and adds vision and computer use, but the weights, licence and full specs are not public yet.
Tencent released Hy4 on 4 November 2026, a flagship model it says now ranks among the world’s leading open-source AI systems. The launch adds vision and computer use for the first time, a 1 million token context window and API access through Tencent Cloud, OpenRouter, WorkBuddy and CodeBuddy.
The catch is in the first paragraph of Tencent’s announcement. The company says it “plans to release model weights and system cards in the coming days.” On launch day there was no public download, no licence and no published parameter count for the full model.
Independent evaluator Vals AI had already tested Hy4 and labels its weights “Open” on the model page, a shorthand for the intended release rather than a file you can download.
Quick answer: Tencent Hy4 is a real API model with vision, computer use and a 1 million token context, and Vals AI ranks it fourth among open-weight models. It is not yet open weight: there are no weights, licence, system card or full parameter count. Vals measured Terminal-Bench 4.0 at 39.39 percent, not the 44.4 percent circulating online, and Hy4 costs more per completed test than DeepSeek V4.1 Flash while scoring slightly lower.
What Tencent actually launched
Hy4 succeeds Hy4 Preview, which Tencent open-sourced in August as a 770 billion parameter mixture-of-experts model with 49 billion parameters active per token. The full release does not repeat those numbers.
Tencent confirms a capability jump. Hy4 handles images, charts and mathematical visual reasoning, and it can operate command-line and graphical interfaces. Tencent says it trained the model on professional software including Blender and CAD.
The company also published internal results on office work, financial analysis, frontend and game development, and scientific research. Those are Tencent-run evaluations with no confidence intervals.
The open-weight claim is ahead of the files
Tencent’s headline says Hy4 joins “the world’s leading open-source AI models.” The operative sentence is that weights and system cards will arrive “in the coming days.”
That is a promise, not a release. There is no licence to read, no config to inspect and no way to check whether the safety exclusions Tencent describes survived into the final checkpoint.
The missing system card matters because Tencent makes specific safety claims. It says offensive cyber material was excluded from training, and that internal red teams found no significant offensive capability in the areas they tested. Those claims deserve scrutiny, but there is nothing to scrutinise yet.
The licence is also unresolved. Hy4 Preview shipped under Apache 2.0. Tencent has not said whether the full model will use the same terms or a more restrictive community licence with revenue limits.
Vals puts Hy4 fourth among open models, not first
Vals AI evaluated Hy4 on its index of professional tasks and ranked it 13th of 33 overall, with 59.35 percent accuracy at $1.72 per test. Among open-weight models it sits fourth.
Kimi K3.1 Pro leads at 62.52 percent, followed by DeepSeek V4.1 Flash 1017 at 61.26 percent and GLM 5.3 at 59.73 percent.
Cost is where the launch pitch weakens. Hy4 costs $1.72 per completed test. DeepSeek V4.1 Flash costs $0.657 and scores 1.91 points higher.
Tencent instead compares Hy4 with GLM 5.3, saying it matches GLM at roughly 20 percent of the cost per task. That is true against GLM at $8.525 per test. It is not true against the cheaper DeepSeek model.
| Model | Vals Index | Rank overall | Rank among open weights | Cost per test | List price per 1M tokens | Context |
|---|---|---|---|---|---|---|
| Kimi K3.1 Pro | 62.52% | 9 of 33 | 1 | $6.44 | $2.00 in / $10.00 out | 1M |
| DeepSeek V4.1 Flash 1017 | 61.26% | 10 of 33 | 2 | $0.657 | $0.30 in / $1.20 out | 1M |
| GLM 5.3 | 59.73% | 12 of 33 | 3 | $8.525 | $1.40 in / $4.40 out | 1M |
| Tencent Hy4 | 59.35% | 13 of 33 | 4 | $1.72 | About $0.90 in / $2.70 out | 1M |
Source: Vals AI model pages, accessed 5 November 2026. Cost per test is total spend to finish its task set, not the sticker token price. Tencent’s list price is approximate and runs with a 50 percent TokenHub discount through 31 December.
Where the benchmark numbers disagree
Terminal-Bench 4.0 is the cleanest example of why launch-week numbers need a second source. Vals measured Hy4 at 39.39 percent, up from 8.08 percent for the preview and second among open-weight models behind Kimi K3.1 Pro at 63.64 percent.
A widely shared launch post from the account Mia AI Lab put the same score at 44.4 percent. Tencent’s press release publishes no Terminal-Bench figure. The 39.39 percent number comes from a reproducible harness, so it is the one to use.
But the same evaluation found a failure mode that matters for agents. Vals said about 21 tasks crashed because Hy4 wrote broken tool calls that Tencent’s API later rejected, and another 18 ran out of time. On the Legal Research Bench it scored 41.35 percent.
👋Hy4 is here, new with vision.
Firmly among the leading open models.
More productive everywhere: finance, frontend, game, officework, research…
Built responsibly, with cyber/biosecurity in mind.
Try it out and let us know!
🔗：https://t.co/ZzL4DtgH3h pic.twitter.com/xceRxN6ueT— Tencent Hy (@TencentHunyuan) November 4, 2026
The finance result is real but narrow
Tencent’s strongest claim is a blind financial evaluation run with Tencent Financial Technology. It used 213 professionals, including 153 from external institutions, and 579 real tasks.
Hy4 scored 3.04, ahead of DeepSeek V4.1 Flash at 2.97 and both Kimi K3.1 Pro and GLM 5.3 at 2.68. A 0.07 lead over DeepSeek is within the range where task selection and judging style can decide the winner.
Tencent also says Hy4 cut its hallucination rate from 12.95 percent to 7.2 percent, and errors from unsupported assumptions from 23.1 percent to 5.6 percent. Those are large improvements on Tencent’s own tasks.
Confirmed versus not confirmed
| Claim | Status | Evidence |
|---|---|---|
| Hy4 launched on 4 November 2026 with a 1M token context | Confirmed | Tencent press release and Vals model page |
| Hy4 adds vision and computer use | Confirmed | Tencent press release |
| Hy4 ranks fourth among open-weight models | Confirmed | Vals AI evaluation, 59.35 percent |
| Hy4 weights are downloadable now | Not confirmed | Tencent says “in the coming days”; no public files or licence |
| Full Hy4 has 770B total and 49B active parameters | Not confirmed | Those are Hy4 Preview figures; Tencent did not repeat them |
| Terminal-Bench 4.0 score is 44.4 percent | Not confirmed | Vals measured 39.39 percent; Tencent published no figure |
| Hy4 will ship under Apache 2.0 | Not confirmed | Hy4 Preview used Apache 2.0; Tencent has not stated the full model licence |
What the price actually looks like
Tencent quotes roughly $0.90 per million input tokens and $2.70 per million output tokens. OpenRouter listed Hy4 at $0.417 input and $1.251 output on 5 November, about half that list price.
The older Hy4 Preview endpoint still shows $0.834 and $2.501, close to the numbers many launch posts quoted as Hy4’s price. Tencent is running a 50 percent TokenHub discount through 31 December, and OpenRouter is passing through a discounted rate.
Tencent's Hy4 debuts at #4 among open-weight models on the Vals Index, behind Kimi K3.1 Pro, DeepSeek V4.1 Flash, and GLM-5.3.
It nearly matches GLM-5.3's performance at about a fifth of the cost per task. It's also a substantial improvement over Hy4 Preview, at about the same… pic.twitter.com/lmKGGar8Vn
— Vals AI (@ValsAI) November 4, 2026
What it means for Indian developers
The API is available now, and that is the part Indian teams can use. A 1 million token context with vision and computer use is useful for document-heavy work such as insurance claims, contract review and back-office automation.
The open-weight promise is the part to treat carefully. Indian companies often want local weights for data residency, cost control or air-gapped deployment. Until Tencent publishes the checkpoint and the licence, none of those plans can be tested.
A sensible sequence is to benchmark Hy4 through OpenRouter against DeepSeek V4.1 Flash on your own tasks, then wait for the weights before committing to self-hosting. Measure cost per completed task, not cost per token. A cheap token price can still produce an expensive workflow when tool calls fail and retries pile up.
Procurement also matters. If the final licence carries revenue or user limits, the economics of a customer-facing product change, as Tencent’s peers in Europe have shown.
Frequently asked questions
Is Tencent Hy4 open source?
Not yet. Tencent says the weights and system cards arrive “in the coming days.” As of 5 November 2026 there is no public download, licence or full parameter count. Hy4 Preview is open weight under Apache 2.0.
How good is Hy4 compared with Kimi K3.1 Pro?
Vals scores Hy4 at 59.35 percent and Kimi K3.1 Pro at 62.52 percent. Hy4 is fourth among open-weight models; Kimi is first, at $6.44 per completed test against Hy4’s $1.72.
What is Hy4’s Terminal-Bench 4.0 score?
Vals measured 39.39 percent, second among open-weight models behind Kimi at 63.64 percent. A 44.4 percent figure circulated online, but Tencent published no Terminal-Bench number and the higher figure does not match the harness.
How much does Hy4 cost?
Tencent lists about $0.90 per million input tokens and $2.70 per million output tokens, with a 50 percent TokenHub discount through 31 December 2026. OpenRouter showed $0.417 input and $1.251 output on 5 November.
Can I run Hy4 on my own servers?
Not today. The weights are not public. The preview shows the family can run as a mixture-of-experts checkpoint, but the full model’s size, licence and hardware needs are unconfirmed. Teams that need self-hosting now should look at Hy3 or DeepSeek.
Sources
Tencent, “Tencent Launches Hy4, Joining the Ranks of the World’s Leading Open-Source AI Models,” 4 November 2026. tencent.com
Vals AI, “Hy4” model page and evaluation note, 4 November 2026. vals.ai
Vals AI, “Kimi K3.1 Pro” model page, 24 October 2026. vals.ai
Vals AI, “DeepSeek V4.1 Flash 1017” and “GLM 5.3” model pages, accessed 5 November 2026. vals.ai
OpenRouter, “Tencent: Hy4” model listing and API pricing, accessed 5 November 2026. openrouter.ai
Hugging Face, “tencent/Hy4-preview” model card, 28 August 2026. huggingface.co
Tencent Hunyuan, launch post on X, 4 November 2026. x.com
Vals AI, evaluation thread on X, 4 November 2026. x.com
AImagazine, “Tencent Hy3: Run a 295B Flagship MoE Model on a Single GPU with GGUF and llama.cpp,” 14 July 2026. aimagazine.blog
Updated 5 November 2026. Tencent had not published the Hy4 weights, licence or system card at the time of writing.