Tencent Open-Sources a 770B Model That Helped Build Itself

Tencent gave away a 770-billion-parameter frontier model on Friday, and the most arresting claim in the launch notes was not a benchmark score. It was that the model had helped build itself.

On August 28, Tencent's Hunyuan team released and open-sourced Hy4 preview, a mixture-of-experts language model with 770 billion total parameters, 49 billion activated per token, and a context window exceeding one million tokens. The weights went live on Hugging Face the same day under an Apache 2.0 license, in both BF16 and FP8 variants, with mirrors on ModelScope, GitCode and CNB for mainland China developers. API pricing landed at $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache-read tokens — cheap enough that a 770B model now costs less to query than most mid-tier closed models did a year ago.

The architecture is a 78-layer backbone. The first layer uses a conventional dense feed-forward network; the remaining 77 replace it with MoE blocks, each holding 256 routed experts plus one shared expert, with every token activating the top eight routed experts alongside the shared one. A native multi-token prediction layer supports speculative decoding. It is an infrastructure-heavy model: open weights that most teams will still rent rather than run.

The evidence Tencent chose to lead with

Rather than opening on a leaderboard, Tencent led with a blind human evaluation. 163 Tencent experts scored model outputs across 203 engineering tasks. Hy4 preview averaged 2.99 out of 4.00, against 2.92 for Z.ai's GLM-5.3 and 2.94 for Moonshot's Kimi K3. The margins are thinner than the averages imply: Hy4 won 46.8% of head-to-heads against GLM-5.3 while losing 40.4%, and won 51.2% against Kimi K3 while losing 40.9%.

Public benchmarks trace the same narrow shape. On SWE-bench Multilingual, Hy4 preview scored 82.9, ahead of GLM-5.3 at 81.3, Kimi K3 at 80.8 and DeepSeek V4 Pro at 77.3. On SWE-bench Pro it reached 65.7 against 64.6 and 63.3 — respectable, though Claude Opus 5 still owns that category at 79.2. Coding-agent maker Cline reported Hy4 leading its own SWE-bench Pro runs and called it the biggest generational leap the company has measured to date.

The stranger claim sits further down the release. Tencent says Hy4 preview participated in its own development, proposing approaches for training methods, data strategies, evaluation frameworks and low-level operators, then running experiments and iterating on the results, with the resulting code and logs feeding the next round. It also analyzed bottlenecks in its own inference stack and performed multiple rounds of operator fusion and communication optimization, lifting end-to-end throughput 31.8% over baseline. Tencent describes this as an early-stage recursive self-improvement loop — language most Western labs reserve for safety papers, deployed here as a product bullet.

Why it matters

Hy4 did not arrive alone. Between August 20 and 28, five labs shipped open-weight models with million-token context windows: Z.ai's GLM-5.3-Flash, Alibaba's Qwen3.8-Flash, MiniMax M3, DeepSeek's V4-Flash-Vision-Exp, and now Hy4. Requesty's Thibault Jaigu, tracking the wave, found model-launch chatter in his firm's social listening corpus jumped from 506 mentions in the week of August 15 to 978 the following week — a 1.93x spike inside nine days.

His read of it is the sharpest available. “Nobody in that list claimed a decisive capability lead,” Jaigu wrote. “What changed is the price attached to the capability.” Three things that commanded a premium eighteen months ago — million-token context, native multimodality, permissively licensed open weights — are now table stakes in the cheap tier. Jaigu's conclusion: “the model is now the cheap part, and the expensive part is everything around it: effort settings, cache hit rate, provider choice and retry behaviour.”

That reframing matters more for Tencent than for its rivals, because Tencent is not selling tokens so much as seats. Hy4 preview shipped simultaneously into WorkBuddy, CodeBuddy, Yuanbao and ima, free on the first two for a fortnight, with free Hy3 access extended through September 30. The model was co-designed with those products and trained on data built with Tencent's own software engineering, gaming, finance and security specialists. Open weights are the marketing; the coding agent is the business.

Behind it is Yao Shunyu, the former OpenAI researcher Tencent recruited to run Hunyuan LLM and AI Infra, who has spent 2026 rebuilding the company's pre-training and reinforcement learning infrastructure. He has consistently refused the framing that Tencent is behind. “AI is a long-term game, with the second half of the race just starting,” he said at Tencent's Cloud AI Industry Applications Summit in June, comparing the moment to personal computing in the 1970s. He also rejected the idea that ChatGPT and Claude Code would remain the only durable AI products: “That would be a very bleak world. Instead, a steady stream of new opportunities will inevitably emerge.” Tencent has said it aims to more than double AI investment past $5 billion this year.

The caveats are real, and Tencent flags some of them itself: the model can spend longer than necessary reasoning through complex tasks and can over-verify work that is already correct — both of which inflate latency and token spend inside agent loops. Writing for MyClaw the day of the launch, Nathan Cole put it plainly: Hy4's “preview status, narrow benchmark margins, and official latency warnings rule out blind adoption.”

What to watch

Three things. First, independent replication — Tencent's 74.1 on Toolathlon-Verified and its blind-eval win rates are vendor scorecards until a third-party harness reproduces them. Second, provider variance: open weights get served at wildly different quantizations and prices, and one developer measured an 11x spread between the cheapest and most expensive host of a single open model this month. Third, and most consequential, the recursive claim. If Hy4's contribution to its own training pipeline holds up in the next release, Tencent will have an argument that no benchmark table can make. Hunyuan says the next batch of Hy4 models rolls out soon.