AI-generated concept illustration — not a screenshot, product photo, or image from Tencent.
On November 4, 2026, Tencent released the full version of Hy4, and the Tencent Hy4 full release is one of the most consequential AI launches of the year. Ten weeks after the August preview introduced a 770-billion-parameter open-weight model built for agents, the finished Hy4 adds two capabilities the preview never had: genuine visual understanding and computer use — the ability to operate both command-line and graphical interfaces, including professional software such as Blender and CAD tools. Tencent describes the launch as its first entry into the top group of open models, a claim that is mostly supported by Tencent's own testing so far. This guide covers everything verified about the release: the architecture, the new vision and computer-use features, the pricing, how to access and use the model, how it compares with rivals like GPT-6 Astra, and what is still unconfirmed.
1. What the Tencent Hy4 full release actually is
Hy4 is Tencent Hunyuan team's flagship foundation model, and the November 2026 full release turns the August text-only preview into a multimodal, computer-using AI system. The preview, released August 28, 2026, was a pure language model: text in, text out, optimized for coding agents and long tool-use chains. The full release keeps that agentic DNA and adds visual reasoning — fine-grained image processing, complex chart analysis, and mathematical reasoning over visuals — plus computer use across command-line and graphical interfaces. Tencent also reports better instruction following, stronger long-context comprehension, and improved performance on long-horizon agent tasks, meaning jobs that run over many steps and actions. One important caveat runs through this entire guide: unlike the preview, which shipped with downloadable weights, a model card, and an Apache 2.0 license on day one, the full release arrived with no downloadable weights, no model card, and no stated license. Tencent says it plans to release the weights and system cards in the coming days, but until that happens, the full Hy4 is available only through hosted APIs and Tencent's own applications. Wherever a claim below comes from Tencent alone rather than an independent source, this article says so.
2. From Hy4 preview to full Hy4: everything that changed
The full release adds vision and computer use while keeping the same reported size and the same API price as the preview. The table below separates what is confirmed from what is merely carried over from preview reporting, because the launch announcement itself did not restate the parameter count or context length — the figures in circulation come from the August preview, and one tracking service notes explicitly that Tencent did not restate them.
| Spec | Hy4 Preview (August 2026) | Hy4 full release (November 2026) |
|---|
| Total parameters | 770B | 770B reported; not restated by Tencent at launch |
| Active parameters per token | 49B | 49B figure in circulation from the preview |
| Architecture | Mixture-of-experts: 256 routed experts plus 1 shared expert per layer, top-8 routed per token | No new architecture description published |
| Context window | 1,048,576 tokens | Preview figure; not restated |
| Max output | 64,000 tokens | Not restated; assume preview figure until confirmed |
| Input modalities | Text only | Text and images |
| Computer use | No; built for fixed tool schemas | Yes: command-line and graphical interfaces |
| Vision | None | Fine-grained image processing, chart analysis, visual math reasoning |
| License | Apache 2.0 | Unstated for the full release |
| Weights | Published on day one (Hugging Face, ModelScope, GitCode, CNB) | Not yet published; promised in the coming days |
| Model card | Published with the preview | Not yet; system cards promised |
| API price | $0.834 per million input tokens, $2.501 per million output tokens, $0.042 cached input | Unchanged, per API listing and routing-service reports |
The practical upshot is straightforward. If you were already calling the preview through an API, the full model gives you vision and desktop control at the same listed price. If you planned to download and self-host the weights, you should wait: the preview's Apache 2.0 weights are downloadable, but the full model's weights, license, and model card do not exist publicly yet.
3. The 770B mixture-of-experts architecture, explained simply
Hy4 is a mixture-of-experts model, which means it stores 770 billion parameters but only activates about 49 billion of them for any single token. Think of it as a hospital with 256 specialist doctors on staff: each incoming patient (token) is routed to the 8 specialists most relevant to the case, plus one general practitioner (the shared expert) who sees everything. That routing happens across 78 layers, with the first layer kept dense. The result is a model with the knowledge capacity of a 770B system and the per-token compute cost of a much smaller one — roughly 49B active parameters. Two engineering details from the preview deserve attention. First, Hy4 uses a gated sparse-attention mechanism (built on DeepSeek's sparse-attention ideas) that lets it handle the full million-token context efficiently instead of drowning in it. Second, it ships with a multi-token-prediction layer of about 10 billion parameters dedicated to speculative decoding — predicting several tokens at once so generation runs faster. On routing services, the preview has been measured at about 54.6 output tokens per second with a median time to first token of around 6.26 seconds, and the full release is reported to have a shorter chain of thought than the preview, which would reduce per-task token consumption. That last claim matters economically: because the model bills per token, fewer reasoning tokens per finished task means cheaper tasks even at an unchanged per-token rate — though Tencent has published no before-and-after token counts, so treat the saving as plausible but unmeasured.
4. Vision, part one: fine-grained image understanding
The headline vision capability is fine-grained image processing — Hy4 can find small, specific details inside very large images. Tencent's demonstration uses a 2.8-gigapixel aerial photograph in which the model locates a two-metre roof component, a striking illustration of detail resolution, though independent analysts correctly note that it is a vendor illustration with no published visual benchmark attached. What matters for real users is the combination of perception with tool calls: Hy4 does not just describe what it sees, it can act on it, chaining visual observations into multi-step work. A few honest cautions belong here. Tencent has not described the vision encoder at all — not its size, not its training data, and not whether it is a separate component — so the architecture behind the vision capability is currently a black box. There are also no independent vision benchmarks yet, and Tencent itself acknowledges that Hy4 still lags the leading closed-source systems in advanced visual question answering. For everyday uses such as reading screenshots, identifying UI elements, inspecting product photos, or describing complex scenes, the capability is real and available through the API today; for research-grade visual reasoning, wait for third-party tests before trusting it with anything critical.
5. Vision, part two: charts, documents, and math over visuals
Hy4 combines visual perception with mathematical reasoning, so it can read a complex chart or diagram and actually work with the numbers in it. According to Tencent, the model handles complex chart analysis and mathematical reasoning over visuals — the kind of task where an analyst pastes a dense financial chart or a scientific figure and asks for the trend, the outliers, or the computed result. This is where the million-token context and the vision capability reinforce each other: long documents full of figures can be ingested whole, and the model can cross-reference what it reads in text with what it sees in images. Tencent validated the science side on TenureTrail, a set of 123 executable tasks across seven disciplines, though this is an internal suite rather than a public benchmark. Practical advice: when you feed Hy4 a chart, state the units and the question precisely, because vision models are strongest when the task is well-specified — ask for the exact figure or comparison you need rather than a vague summary. And remember the preview's documented weakness: it tended to over-verify its own work, reasoning longer than necessary on hard tasks. The full release is said to have a shorter chain of thought, but until independent measurements exist, budget extra tokens for hard analytical work.
AI-generated concept illustration — not a screenshot, product photo, or image from Tencent.
6. Computer use: the command line, now under AI control
Hy4 can operate a command-line interface directly — running shell commands, navigating file systems, and driving terminal-based workflows as an agent. This is a step beyond the preview, which was built for fixed tool schemas: a fixed list of functions the developer defines in advance. Computer use means the model interacts with the machine the way a human operator would, issuing commands and reading their output in a loop. The preview already showed where this was heading — it scored 85.4 on Terminal-Bench 2.1, a benchmark that grades models on driving a real terminal, a figure within a few points of independently measured closed models. With the full release, that terminal-driving ability becomes a first-class feature rather than an emergent trick. The safety implications are obvious and Tencent has responded to them: the company says it filtered training material on exploit development, reverse engineering, and penetration testing out of Hy4's training, focusing instead on defensive cybersecurity uses to limit how much the model can help with attacks. Independent analysts add the practical guidance that still applies: define tool permissions, set timeouts and retry limits, and require a human approval step before an agent with shell access takes irreversible actions. Never give any computer-use agent unrestricted shell access on a machine you care about — sandbox it first, exactly as Tencent does with its own Agent Runtime.
7. Computer use: graphical interfaces, Blender, and CAD
The more surprising half of computer use is graphical: Hy4 can interact with graphical user interfaces, and Tencent says it received specialized training on professional software such as Blender and CAD tools. One routing service called this the most significant addition in the release — the first time the Hy line claims agency over a desktop rather than over a tool schema. In practice, this means the model can see a screen, locate buttons, menus, and canvases, and perform multi-step workflows inside real applications. Tencent runs this in a sandbox built on its cloud Agent Runtime with several hundred applications installed, which tells you how seriously the company takes the containment question. The professional-software angle is the commercial tell: 3D modeling in Blender and design work in CAD are long, fiddly, multi-hundred-click workflows where an AI assistant that can actually drive the interface is worth far more than a chatbot that describes the steps. Two caveats. First, this is brand new — there are no public computer-use benchmarks for Hy4 yet, so every claim about reliability comes from Tencent. Second, graphical agents are inherently brittle: a UI update, a resized window, or an unexpected dialog can derail them, so treat early computer-use work as supervised automation with a human watching, not as a fire-and-forget employee.
8. Better instruction following and long-context comprehension
Tencent reports that the full Hy4 follows instructions more reliably and understands long contexts better than the preview did. These sound like incremental improvements, but for agent workloads they are the difference between a demo and a deployment. Instruction following is what keeps a 40-step task on track at step 31: the model must remember the original goal, respect constraints stated early in the conversation, and not drift into plausible-but-wrong side quests. Long-context comprehension matters because Hy4's context window is about a million tokens — roughly an entire codebase, a full legal discovery set, or months of chat history — and a big window is worthless if the model loses track of details buried in the middle of it. The preview already offered three reasoning levels (high, low, and none, with high as default) plus a documented sampling recommendation of temperature 0.9 and top_p 1.0; the full release keeps that control while reportedly tightening the underlying behavior. Tencent's own WorkBuddy evaluation, which grades the model across workplace scenarios, gave Hy4 an average score of 83.18 with a 96.15 percent pass rate, ranking first in nine of twelve scenarios in the mainland Chinese market. As always, that is a vendor-run number — useful as a signal, not as proof.
9. Long-horizon agent tasks: what multi-step work looks like now
Long-horizon agent tasks — jobs that run over many actions, tools, and decisions — are the workload Tencent built Hy4 for, and the full release pushes this further than the preview. The preview's resume was already strong here: 64.3 on DeepSWE (up from 28.0 on its predecessor Hy3), 74.1 on Toolathlon-Verified for tool calling, and 37.1 pass@1 on APEX-Agents, all vendor-reported. The full release adds vision and computer use to that foundation, which unlocks genuinely new task shapes: an agent that reads a dashboard screenshot, runs a database query in a terminal, edits a spreadsheet in a GUI, and files the result — all in one session. Tencent reports a 12.03 percentage-point gain over the preview on iRaB, an internal investment-research agent benchmark, plus gains of 19.1 percent on GameDev and 9.8 percent on GameLogic suites. Independent analysts note these are internal suites with no published item counts, so the honest reading is directional: the full model is better at agentic work than the preview, and the size of the gap awaits independent measurement. If you want to understand the broader shift toward agents that act rather than just chat, our guide to how to use AI agents in 2026 covers the patterns, and our report on AI agents breaking out of their sandboxes covers the safety debate.
10. The known issues from the preview — and what Tencent says it fixed
The preview's own model card admitted two flaws: the model reasoned longer than necessary on hard tasks and over-verified its own answers — and the full release claims to fix the first one. On September 7, 2026, Tencent pushed an optimization into the live preview build that it said cuts reasoning turns and per-task token consumption on complex work; the full release is described as having a shorter chain of thought as standard. Because Hy4 bills per token, shorter reasoning directly lowers the cost of each completed task — but Tencent published no before-and-after token counts, so the efficiency gain cannot be measured yet, only tested on your own workloads. The hallucination numbers are messier: Tencent's announcement says the hallucination rate fell from 12.95 percent in the preview to 7.2 percent, while one routing service reports the figures as 15.76 percent down to 5.45 percent — the sources agree on the direction and on a related figure (errors from unsupported assumptions falling from 23.1 percent to 5.6 percent), but the methodology is unpublished and the headline numbers do not match. The takeaway for users: expect a more economical, more accurate model than the preview, but verify on your own tasks rather than trusting any single percentage. Tencent also acknowledges remaining gaps frankly — Hy4 still lags the leading closed-source systems in research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering.
11. Open weights: what is released, what is promised, and the license question
As of this writing, the full Hy4's weights have not been released — Tencent only says it plans to publish the weights and system cards in the coming days. This is the single biggest difference between the two launches. The August preview arrived with everything an open-source release needs: downloadable weights in BF16 and FP8 on Hugging Face, ModelScope, GitCode, and CNB, a published model card, Docker images for vLLM and SGLang with an OpenAI-compatible API, and an Apache 2.0 license. The full release has none of that yet, and the license for the full release is unstated — the Apache 2.0 terms applied to the preview, and it would be a mistake to assume they automatically carry over. Why the delay? Tencent ties the planned release to a focus on cybersecurity and biological safety, and the company says it filtered cyber-offensive training material out of the model; a cautious staged release is consistent with that posture, and Beijing has published draft rules about the terms on which very large models may be released openly, which may also be a factor. The practical consequence is clear: if self-hosting is your plan, the preview weights remain your only downloadable option today. If you can work through an API, the full model is already live. Watch for the weight drop and read the license file before assuming anything — the license determines whether you can use the model commercially, fine-tune it, and redistribute it.
12. Deployment math: VRAM, quantization, and Docker images for the preview weights
Self-hosting a 770B model is a serious infrastructure commitment: even the 4-bit quantized preview checkpoint is about 435 GiB, which needs a 512GB machine. The full checkpoint in BF16 is far larger, which is why most teams will use the FP8 build or quantized versions served through vLLM or SGLang — Tencent publishes Docker images for both, with an OpenAI-compatible API so existing code works with minimal changes. A few rules of thumb for planning. First, hosting requires storing the full checkpoint even though only 49B parameters activate per token; sparsity saves compute, not disk. Second, the multi-token-prediction layer exists specifically to speed up speculative decoding, so serving stacks that support it will get meaningfully better throughput. Third, the 64,000-token output ceiling from the preview is the spec that decides deployments: long generated files or multi-file patches can hit it and truncate silently, so pipelines need detection and retry logic — one routing service's recommended pattern is to run the loop on the cheap model and escalate to a larger-output model when truncation is detected. None of this applies to the full release yet, because there are no weights to download; treat this section as the preview's proven deployment recipe and expect an updated version when the full weights land.
13. API pricing: the cheapest flagship-class model in its tier
Hy4's API price did not move with the full release: $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens — roughly a twelfth of GPT-6 Astra's input price and a twentieth of its output price. A ZenMux listing for the TencentCloud provider shows the full Hy4 entry at exactly the preview's rates, and routing services describe the two as the same family, the same size, and the same price. The cached-input figure is the sleeper advantage: agent loops re-send the same system prompt and conversation prefix on every turn, so the cheapest tokens in a long loop are 24 times cheaper than Astra's $1.00 cached rate. On a typical 4:1 input-to-output agent loop, independent analysis puts Hy4 at roughly $1.17 per million tokens against roughly $18.00 for GPT-6 Astra — and a 100-million-token month at roughly $117 versus roughly $1,800. For the wider context of collapsing inference prices, see our complete guide to the 2026 AI price war, which tracks verified pricing across every major model.
| Model | Input (per 1M tokens) | Output (per 1M) | Cached input (per 1M) | Source basis |
|---|
| Tencent Hy4 / Hy4 Preview | $0.834 | $2.501 | $0.042 | Provider listings; unchanged at full release |
| GPT-6 Astra | $10.00 | $50.00 | $1.00 | GA pricing since 3 Sep 2026; whole-request reprice above 272K input tokens |
| Gemini 4 Carbon | $0.85 | $5.00 | $0.15 | Enterprise Cloud pricing |
| GLM 5.3 | $2.00 | $10.00 | $1.00 | Via Zhipu, Nov 2026 |
| Kimi K3 | Not in verified sources | $15.00 | Not in verified sources | Output figure from November 2026 price tracking |
| DeepSeek V4 Pro | Not in verified sources | Under $2.00 off-peak | Not in verified sources | Off-peak output figure from November 2026 price tracking |
Two notes on reading this table. First, per-token price is not per-task price: a model that reasons concisely can beat a cheaper-per-token model on total cost, which is why Tencent's claimed shorter chain of thought matters. Second, prices move fast in this market — verify the current rate on the provider's page before committing volume.
14. Where to access Hy4 right now
The full Hy4 is live through API providers today, while the preview's open weights remain the only downloadable option. Your access route depends on what you need: API access for the full model, or weights for self-hosting the preview.
| Channel | What you get | Notes |
|---|
| Tencent Cloud TokenHub | Full Hy4 API | Tencent's own endpoint; listed at $0.834 / $2.501 per 1M tokens |
| OpenRouter | Hy4 Preview API (tencent/hy4-preview); full-model routing expected | One OpenAI-compatible endpoint; list rates passed through with no markup reported |
| ZenMux (TencentCloud provider) | Full Hy4 API listing (tencent/hy4) | Shows the same $0.834 / $2.501 pricing as the preview |
| Yuanbao | Full Hy4 in Tencent's consumer AI app | Easiest free-form way to try vision and chat features |
| ima | Full Hy4 in Tencent's AI workspace app | Document- and knowledge-focused use |
| WorkBuddy | Full Hy4 in Tencent's workplace AI, integrated with WeCom | Reaches chats, emails, documents, calendars, meetings; team Projects feature upcoming |
| CodeBuddy | Hy4 for coding workflows | Terminal and code-agent use cases |
| Hugging Face / ModelScope / GitCode / CNB | Hy4 Preview weights (Apache 2.0) | BF16 and FP8 builds; full-release weights not yet published |
| Self-host via Docker (vLLM / SGLang) | Hy4 Preview serving | OpenAI-compatible API; 4-bit quant ≈435 GiB, needs a 512GB machine |
If you are choosing between routes, the decision tree is simple: try the model in Yuanbao or ima for free-form exploration, use TokenHub or a router for API volume, and only consider self-hosting if you need data residency or custom fine-tuning — and even then, wait for the full-release weights.
15. Using Hy4 inside Yuanbao, ima, WorkBuddy, and CodeBuddy
Tencent built the full release around its own applications, and each one targets a different kind of work. Yuanbao is the consumer front door: the fastest way to test the new vision features is to drop an image, a chart, or a screenshot into a Yuanbao chat and ask Hy4 to analyze it. ima is the knowledge-workspace play — long documents, research material, and the million-token context doing what it was built for. WorkBuddy is the enterprise story and the biggest product news in the release: Hy4 powers a broad WorkBuddy update that integrates with WeCom, Tencent's workplace messaging app, reaching into chats, emails, documents, calendars, and meetings, with team collaboration features arriving in an upcoming Projects capability. CodeBuddy is the developer surface for coding agents and terminal workflows. The strategic logic, as described by Nikkei Asia's coverage of Tencent's broader shift, is a move from proprietary to open-source models to pull customers toward Tencent's cloud services — the model is the magnet, the cloud is the business. For everyday users, the practical advice is to start in Yuanbao or ima: no API keys, no infrastructure, and immediate access to the vision and computer-use features as Tencent rolls them out across its apps.
16. Developer quickstart: OpenAI-compatible API, OpenRouter, and TokenHub
Because Hy4 exposes an OpenAI-compatible API, most developers can switch to it by changing a base URL and an API key, not by rewriting code. The standard path is Tencent Cloud TokenHub for direct access, or a routing service such as OpenRouter or ZenMux if you want one endpoint for multiple models — the preview is catalogued as tencent/hy4-preview and the full model as tencent/hy4, with provider list rates passed through. A few integration tips from the verified record. Set your expectations on output length: the preview's 64,000-token ceiling is the binding constraint, so build truncation detection and retry logic into agent loops rather than assuming every response completes. Use cached input aggressively: at $0.042 per million cached tokens, long agent loops with repeated system prompts are where Hy4's pricing compounds fastest. Start new projects against the preview's documented sampling settings (temperature 0.9, top_p 1.0) and the three reasoning levels, then re-tune when the full-release model card arrives. And keep a second provider warm: routing data shows even premium routes fail about one request in ten over a rolling week, while the Hy4 route has measured around a 2.8 percent error rate — automatic failover is cheap insurance for long agent runs that would otherwise die with their accumulated context.
17. Prompting Hy4 well: practical tips for vision and computer-use tasks
Hy4 rewards precise, verifiable instructions — and punishes vague ones — especially in its two new modalities. For vision tasks, specify exactly what you want extracted: name the units, the comparison, or the decision the image should inform, and ask the model to quote the figures it reads before interpreting them, which makes errors checkable. For chart analysis, paste the chart's context (what the axes mean, the time range) rather than assuming the model infers it. For computer-use tasks, the discipline is different: give the agent a clear goal, explicit boundaries (which directories it may touch, which actions need approval), and a defined stop condition, because an agent driving a GUI can wander. Break long desktop workflows into stages with human checkpoints between them — for example, let it set up the Blender scene, review the result yourself, then let it proceed to rendering. Keep tasks idempotent where possible so a retry does not cause damage. And remember the model's documented tendencies: it has historically reasoned longer than necessary and over-verified its answers, so for straightforward tasks, explicitly ask for the direct answer first and the reasoning summary after — you will save tokens and time.
18. Benchmarks: the preview's scores and the full release's missing appendix
The full release shipped with no benchmark appendix — no Terminal-Bench, no SWE-bench, no public coding scores — so the preview's vendor-reported numbers remain the best available stand-in for the family. Here is what Tencent reported for the preview: 85.4 on Terminal-Bench 2.1 (a real-terminal coding benchmark), 64.3 on DeepSWE (up from 28.0 on Hy3), 74.1 on Toolathlon-Verified for tool calling, 37.1 pass@1 on APEX-Agents, roughly 5th place with 1,633 points on the Arena WebDev AutoEval, and a blind human evaluation of 2.99 out of 4.00 across 203 engineering tasks scored by 163 experts — ahead of Kimi K3 at 2.94 and GLM 5.3 at 2.92. For the full release, Tencent substituted its own suites: a finance expert panel (213 professionals, 579 real tasks) scoring Hy4 at 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3; a 12.03 percentage-point gain over the preview on the internal iRaB investment-research benchmark; 19.1 and 9.8 percent gains on internal GameDev and GameLogic suites; and a WorkBuddy evaluation averaging 83.18 with a 96.15 percent pass rate. The honest assessment, echoed by independent analysts: the preview's Terminal-Bench figure sits only about three points below an independently measured closed model, which makes the parity claim plausible but unproven, and the full release's claims need third-party testing on public coding, vision, and computer-use benchmarks before anyone treats them as settled. Coverage elsewhere notes the wider self-benchmarking beat GLM-5.3 and DeepSeek-V4.1-Flash-1017 on all reported tasks with mixed results against Kimi K3.1 Pro — take it with a grain of salt until independent reviewers weigh in.
19. Hy4 vs the rivals: GPT-6 Astra, Gemini 4 Carbon, GLM 5.3, Kimi K3, and LongCat-2.5
Hy4's competitive position is simple to state: it is the cheapest model in the near-frontier tier, with open-weight heritage and new computer-use tricks — but its headline scores are vendor-reported while its main rival's are independently verified. The table below compares what is actually verified for each model.
| Model | Size / license (verified) | Context | Verified price (per 1M) | Headline verified result |
|---|
| Tencent Hy4 (full) | 770B reported, not restated at launch; license unstated | ~1M tokens (preview figure) | $0.834 in / $2.501 out | Finance panel 3.04/4 (Tencent-run); no public coding scores for full release |
| Tencent Hy4 Preview | 770B MoE, 49B active; Apache 2.0 | 1,048,576 tokens | $0.834 in / $2.501 out / $0.042 cached | Terminal-Bench 2.1 85.4 (vendor-reported) |
| GPT-6 Astra | Closed; GA 3 Sep 2026 | 1,050,000 tokens; 128K max output | $10 in / $50 out / $1.00 cached | Terminal-Bench 2.1 88.4 (independently evaluated); 54.7 Intelligence Index |
| Gemini 4 Carbon | Closed | Not in verified sources | Not in verified sources | No verified head-to-head figures in sources checked |
| Z.ai GLM 5.3 | Open weights (GLM family) | Not in verified sources | Not in verified sources | 2.92/4 in Tencent's blind eval; GLM 5.3 Flash holds ~8.5% of tracked gateway tokens |
| Moonshot Kimi K3 / K3.1 Pro | Open weights (Kimi family) | Not in verified sources | Not in verified sources | 2.94/4 in Tencent's blind eval; 2.68/4 in Tencent's finance panel |
| Meituan LongCat-2.5 | 1.6T parameters reported | Not in verified sources | Not in verified sources | See our LongCat-2.5 launch roundup for the full breakdown |
The matchup that matters most is Hy4 versus GPT-6 Astra, because both publish a Terminal-Bench 2.1 figure on the same test: Astra's 88.4 is independently evaluated, Hy4 Preview's 85.4 is vendor-reported, and the three-point gap is well inside what a different harness or scaffold can produce — so the honest reading is plausible parity, not a proven ranking. Astra's advantages are verified scores, twice the output ceiling (128K vs 64K), and image/file input on the preview comparison; Hy4's advantages are roughly 12x cheaper input, 20x cheaper output, open-weight heritage, and now vision plus computer use. Against the Chinese open-weight field — GLM 5.3, Kimi K3, DeepSeek's V4 line — Hy4's finance-panel scores lead Tencent's own tests, but those tests are Tencent's, and the field is moving weekly. For the China-model landscape in full, our Apple Intelligence in China coverage shows how deeply these models are embedding into consumer platforms.
20. Who should use Hy4 today, and who should wait
Use Hy4 through the API today if you run text-heavy agent loops, need vision or computer-use features, or want the cheapest capable tokens in the tier — and wait on the weights if self-hosting is your plan. The ideal early adopters are developers building coding agents and tool-use pipelines (the preview's Terminal-Bench and DeepSWE scores suggest real capability here), teams experimenting with GUI automation in sandboxed environments, analysts working with charts and long documents, and anyone whose workload fits the 64K output ceiling and benefits from $0.042 cached input. You should wait if you need downloadable weights under a clear license — those do not exist yet for the full release — if your compliance team requires independent benchmarks before adoption, or if your outputs routinely exceed 64,000 tokens per response. Enterprises should also note the geopolitical layer honestly: US federal agencies have been ordered to remove DeepSeek products and lawmakers have scrutinized Chinese models, so check your data-residency terms and keep a second provider warm regardless of which model you pick. The roadmap gives everyone something to watch: Hy4 Flash, a lighter version meant to cut task costs, is next, followed by Hy5 — and Tencent's capital expenditure rose 176 percent year over year to 52.8 billion yuan (about $7.9 billion), which signals how heavily the company is betting on this line. The weights and system cards, promised in the coming days, are the event that will decide whether the full Hy4 becomes the open-weights milestone Tencent claims or remains an API-only showcase.
Frequently Asked Questions
Is Tencent Hy4 open source?
Not yet in practice for the full release. Tencent says it plans to release the model weights and system cards in the coming days, but as of this writing no weights, no model card, and no license have been published for the full Hy4. The August preview was released under Apache 2.0 with downloadable weights — do not assume the same license applies to the full release until Tencent states it.
What is new in Hy4 compared with Hy4 Preview?
Vision and computer use are the two headline additions. The preview was text-only; the full release adds fine-grained image understanding, chart analysis, and math reasoning over visuals, plus the ability to operate command-line and graphical interfaces — including specialized training on professional software like Blender and CAD. Tencent also reports better instruction following, long-context comprehension, and long-horizon agent tasks, plus a shorter chain of thought.
How big is Tencent Hy4?
The figures in circulation are the preview's: 770 billion total parameters with 49 billion active per token and a context of about one million tokens. Tencent did not restate the parameter count or context length in the full-release announcement, so treat the 770B/1M figures as reported carry-overs, not freshly confirmed specs. One outlet reports 770 billion for the full model; the architecture details (256 routed experts, top-8 routing, 78 layers) are preview figures.
How much does the Hy4 API cost?
$0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens — unchanged from the preview. That makes it roughly 12 times cheaper than GPT-6 Astra on input and 20 times cheaper on output. On a typical 4:1 input-to-output agent loop, independent analysis puts Hy4 near $1.17 per million tokens versus about $18.00 for Astra.
Can Hy4 really use a computer?
Yes — it can interact with both command-line and graphical user interfaces, a first for the Hy line. Tencent says the model received specialized training on professional software such as Blender and CAD, and runs computer use inside a sandbox built on Tencent Cloud's Agent Runtime with several hundred applications installed. There are no public computer-use benchmarks yet, so treat early use as supervised automation with a human watching, with defined tool permissions, timeouts, and approval steps.
Does Hy4 support images?
Yes — visual understanding is new in the full release. Hy4 handles fine-grained image processing (Tencent's demo finds a two-metre roof component in a 2.8-gigapixel aerial photo), complex chart analysis, and mathematical reasoning over visuals, combining perception with tool calls. Tencent has not described the vision encoder's size or training data, and no independent vision benchmarks exist yet.
How does Hy4 compare with GPT-6 Astra?
Hy4 is dramatically cheaper with plausible capability parity on terminal coding — but only Astra's scores are independently verified. Both have a Terminal-Bench 2.1 figure: Astra 88.4 (independent) versus Hy4 Preview 85.4 (vendor-reported). Astra offers twice the output ceiling (128K vs 64K tokens); Hy4 offers open-weight heritage, 12–20x lower prices, and now vision plus computer use. For text-in, text-out agent loops that fit the output ceiling, Hy4 is the economical choice; for anything needing verified scores or very long outputs, Astra holds the edge.
What are Hy4's known limitations?
Tencent itself acknowledges that Hy4 still lags leading closed-source systems in research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering. The preview's documented flaws — over-long reasoning and over-verification — are said to be improved via a shorter chain of thought, but no before-and-after token counts are published. The full release has no public coding benchmarks, no published weights or license yet, and a 64,000-token output ceiling carried over from the preview that can silently truncate long generations.
When will the Hy4 weights be released?
Tencent says the weights and system cards will arrive in the coming days, but no date is confirmed. Until then, the full model is API-only, and the preview's Apache 2.0 weights remain the only downloadable option. When the weights land, read the license file carefully — the preview's Apache 2.0 terms may or may not carry over, and the license determines commercial use, fine-tuning, and redistribution rights.
Is Hy4 safe to use for business workloads?
Tencent has taken visible safety steps, but standard agent precautions still apply. The company filtered exploit-development, reverse-engineering, and penetration-testing material from training, focusing on defensive cybersecurity uses, and runs computer use in a sandboxed Agent Runtime. For your own deployments: sandbox every agent with shell or GUI access, define tool permissions and timeouts, require human approval for irreversible actions, check data-residency terms, and keep a second provider warm for failover.
Sources
All facts in this article were verified against the sources below on 5 November 2026. Figures attributed to Tencent are vendor-reported unless noted as independently measured.
- Tencent releases full Hy4 AI model that can see images and use a computer — Unicorn API, November 2026 (full-release details, pricing, benchmarks, safety)
- Tencent Hy4 Preview: 770B open-weight AI model built for agents — Unicorn API, August 2026 (preview architecture, weights, known issues)
- GPT-6 Astra vs Tencent Hy4 Preview: Who Checked the Numbers — OrcaRouter, October 2026 (independent pricing, benchmark comparison, route reliability data)
- Tencent Hy4 Preview open-sourced as 770B Apache 2.0 MoE — You.com (preview open-source release details)
- Tencent Hy4 Preview — i-SCOOP (preview specifications and serving details)
- AI model releases, November 2026 — AF.net realtime tracker (release timeline context)
- Model launch signals — Superpower Daily (launch signal tracking)
- The Great AI Price War — AI Tech Hub (rival pricing: Kimi K3 output, DeepSeek V4 Pro off-peak, GPT-6 Astra context)
- LongCat-2.5: Meituan's 1.6-trillion-parameter launch — AI Tech Hub (rival landscape)
The Bottom Line
The Tencent Hy4 full release is a genuine step forward — vision plus computer use at an unchanged, market-leading price — but it is a launch with homework still outstanding. What is verified: a 770B-class MoE model (preview figures, not restated) that now sees images, reads charts, reasons over visuals, and drives command-line and graphical interfaces including Blender and CAD; API pricing of $0.834 input and $2.501 output per million tokens; availability through TokenHub, routers, and Tencent's Yuanbao, ima, WorkBuddy, and CodeBuddy apps; and vendor-reported gains across finance, game-dev, and workplace benchmarks. What is not yet verified: the weights, the license, the model card, the vision encoder's design, and any independent benchmark of the full release — nearly every headline claim still rests on Tencent's word alone. If you build agents, try it through the API this week: the price makes experimentation nearly free, and computer use is the capability most worth testing hands-on. If you self-host or need audited numbers, wait for the weights and the first third-party evaluations. Either way, Hy4 confirms the direction of the entire industry in late 2026: the frontier is getting cheaper, the models are getting hands, and the next competition will be fought on what AI can operate — not just what it can say.
Comments
Post a Comment