On November 4, 2026, Tencent Hunyuan released the full version of Hy4 — just over two months after the August 28 preview, but this is more than a routine iteration: vision understanding and computer use join the Hunyuan product line for the first time, pretraining and post-training scale are significantly expanded, and four productivity lines — office, finance, games and research — launch at once. Tencent's line is that Hy4 "joins the ranks of the strongest open-source models for the first time," yet the same release unusually devotes a whole section to its gaps. That combination of confidence and self-scrutiny is the most notable thing about this launch.
The core change is two "firsts": vision understanding and computer use. On vision, Hy4 handles fine-grained image processing, complex chart analysis and mathematical visual reasoning; in Tencent's demo it located a roughly 2-metre rooftop component in a 2.8-billion-pixel aerial image and correctly read two subplots out of an 80,000-point scatter chart. On computer use, it combines command-line and graphical-interface operation, with specialist training on Blender, CAD and other professional software — not just clicking around, but operating with the logic of the professional tool.
Second, scale and capability across the board: pretraining and post-training scaled up significantly, with instruction following, long-text understanding, long-horizon agents and context utilization all improved. The official appendix shows Terminal-Bench 4.0 rising from the preview's 26.3 to 44.4, DeepSWE from 64.3 to 74.9, APEX-Agents from 53.9 to 75.2. Hallucination rates fell too: Tencent's internal general test dropped from 12.95% to 7.2%, and the search-optimized internal test from 15.76% to 5.45% — but these are Tencent's own test harnesses, and independent measurement is considerably harsher (see the benchmarks article).
Third, a complete story across four productivity lines. Office: the WorkBuddy office evaluation averaged 83.18 with a 96.15% pass rate, No. 1 in China in 9 of 12 scenarios. Finance: in Tencent's FiT crowdsourced test, 213 experts (153 external) scored it 3.04, best among open models. Games: GameDev +19.1%, GameLogic +9.8% (vs the preview). Research: an official demo reconstructed a mouse embryo in 3D — 84 unordered 2D slices, 6.99 million cells, 1.5 hours with no human intervention. All of these are Tencent's own measurements or Tencent-organized evaluations; keep that in mind when reading them.
TokenHub list prices are identical to the preview: ¥6 in, ¥18 out per 1M tokens, ¥0.3 cache hits; international $0.834 / $2.501 / $0.042. A 50% discount runs from launch through Dec 31, 23:59 Beijing time. On channels, TokenHub (model ID hy4) and OpenRouter (tencent/hy4) went live at launch, with Hermes and OpenCode integrated; Tencent's own products — WorkBuddy/CodeBuddy, Yuanbao, ima and the WeCom assistant Dayuan — switched to Hy4. WorkBuddy and CodeBuddy offered a free quota Nov 4–15 — users hit the limits on day one, and Tencent expanded the inference cluster tenfold the next day and raised the quota.
The launch page devotes a section to limitations: in frontier math, long-horizon development, model iteration and high-difficulty visual Q&A, "a significant gap remains" to the best closed models. The official chart backs that up: Hy-FrontierMath T5 at just 13.0 (GPT-6.1 Sol 97.5, Claude Opus 5.5 91.3), Goal2Repo 23.2 (Claude 74.0), Terminal-Bench-Science 19.8 (GPT 60.9). Caixin quoted the Hunyuan team as saying it will "steadily expand pretraining and post-training scale to narrow the gap" — and Hy5 will be larger still.
That candour is more than posture; it also explains Hy4's positioning: not competing for "strongest model," but for a seat among the "strongest open-source models" — No. 1 among open models in office, finance and computer-use scenarios, at a fraction of the price of same-scoring closed models. Artificial Analysis' independent measurement confirms the positioning: index 35, tying Grok 4.7 (xhigh), at about $0.78 per task — roughly one-tenth of the same-scoring Grok.
Nine days after launch (Nov 13), the full-release weights are still unpublished, the license is unpublished, and no system card has appeared. Tencent says "coming soon," but gives no date; a Nov 11 re-check of Hugging Face still found only preview weights. Hy4 Flash is likewise only "coming soon," with no specs, date or price. For users who want local deployment or self-hosting, the only thing usable from this launch so far is the API — and the preview weights are a text-only model without the two headline capabilities of this release.
Reference: Tencent — Tencent Launches Hy4, Joining the Ranks of the World's Leading Open-Source AI Models (rewritten and organized by ModelBrief; facts per the official source)