Frontline Hotspot
Frontline Hotspot

Tencent Open-Sources Hy4 Preview: A 770B Flagship That Helped Train Itself

On August 28, Tencent released and open-sourced its new flagship Hy4 preview (770B total / 49B active MoE, 78 layers): Gated DSA sparse attention + IndexCache cross-layer index reuse + iHC identity Hyper-Connections, with the README openly stating the architecture is "inspired by DeepSeek and GLM". A native MTP layer enables 3-token speculative decoding, context spans 1M tokens, and BF16+FP8 weights ship under Apache 2.0. In Tencent's internal blind eval, 163 experts scored 203 engineering tasks at 2.99/4.00, edging out GLM-5.3 (2.92) and Kimi K3 (2.94, both internal-caliber numbers). The headline is the early loop of recursive self-improvement: the model took part in automating optimization of its own training methods, data strategies, eval frameworks and low-level operators, and autonomously lifted inference end-to-end throughput by 31.8%. OpenRouter snapshot pricing: $0.834 input / $2.501 output per million tokens; free for two weeks on WorkBuddy/CodeBuddy.

Published August 29, 20266 min read
<!-- tencent-hy4-preview-open-source-hotspot | hotspot | Tencent Open-Sources Hy4 Preview: A 770B Flagship That Helped Train Itself -->

On August 27, Tencent Hunyuan announced its next-generation flagship Hy4 preview and open-sourced the weights: 770B total parameters, 49B activated per token, a 1M context window, Apache 2.0 licensing, with BF16 and FP8 weights landing simultaneously on Hugging Face, ModelScope, GitCode, and CNB. The parameter count is just the appetizer - the most striking line in the official announcement is that Hy4 preview participated for the first time in the automated optimization of its own training process, "establishing an early-stage recursive self-improvement loop." A flagship model shipping open-source with the footnote "it helped train itself" is a first for the open-source camp.

Scope note: this article is based on Tencent's official press release (2026-08-28), the Hunyuan GitHub README, and public reporting (as of 2026-08-29). GitHub stars are an API snapshot from 2026-08-29; blind-eval figures are Tencent's internal evaluation methodology; API prices are snapshots and the official pages prevail. Not investment advice.

1. Third Major Release in Six Months: The Preview Playbook Is Working

First, the cadence. This is Tencent Hunyuan's third major model release in six months (per Pandaily). Back in May, Hy3 preview topped OpenRouter's call volume under an anonymous, low-price identity (input as low as $0.066 per million tokens) - the community rushed in first, and the official reveal came later. Hy4 preview follows the same "preview first, collect real-world feedback, then ship the official version" path. The press release puts it directly: "Through a preview-first approach, followed by official releases, Hunyuan continuously incorporates real-world feedback into its research and development process."

The GitHub repository Tencent-Hunyuan/Hy4-preview was created on August 27 and had gathered 245 stars and 10 forks by August 29. For a 770B model, stars are not the point - full weight openness is: both BF16 and FP8 versions are up on four platforms under Apache 2.0, free for commercial use. Tencent also teases that "the next batch of models in the Hy4 series is expected to roll out soon" - the preview is only the opening act.

2. Architecture: 78 MoE Layers, Only 49B Activated per Token

The spec sheet, from the README (backbone only, excluding the MTP layer):

PropertyValue
Total parameters770B
Activated parameters49B
Layers78 (first layer dense FFN + 77 MoE layers)
Expert config256 routed experts + 1 shared expert per layer; top-8 routed + shared activated per token
Hidden size6144
AttentionGated DSA (64 heads), sparse indexer top-k 2048
Residual streams4 (iHC hyper-connections)
Context length1M
Vocabulary120,832

Three technical points deserve a closer look. First, attention: Gated DeepSeek Sparse Attention (Gated DSA, arXiv:2512.02556) with IndexCache (arXiv:2603.12201) for cross-layer sparse index reuse - the README says outright that the design is "inspired by DeepSeek and GLM," naming its peers as teachers in the technical report. Second, the residual pathway: iHC (identity Hyper-Connections) expands inter-layer information flow into 4 residual streams, giving gradients and information multiple routes through 78 layers of depth. Third, a native MTP layer (10B total / 0.7B activated) built for speculative decoding, accepting 3 draft tokens by default at deployment - a key lever for keeping inference costs down on a 770B model.

For anyone deploying it, the threshold number is "all 770B weights resident in VRAM, only 49B computed per token": a small activation saves compute and per-token latency, but the memory bill shrinks not at all. The companion SOP in this batch walks through the official prebuilt-image deployment path, and the companion review lines up the open-source flagships' deployment thresholds in one table.

3. Blind Eval: 163 Tencent Experts Score It 2.99, Slightly Ahead of GLM-5.3 and Kimi K3

Tencent's evaluation methodology is unusual: instead of chasing leaderboards, it had 163 internal experts (software engineers, game developers, finance analysts, security experts) blind-rate outputs on 203 real engineering tasks. Hy4 preview scored 2.99/4.00:

ModelBlind-eval average (out of 4)Win / Tie / Loss
Hy4 preview2.99--
GLM-5.32.9246.8% / 12.8% / 40.4%
Kimi K32.9451.2% / 7.9% / 40.9%

The win rates are solid (46.8% against GLM-5.3, 51.2% against Kimi K3), but ties plus losses also add up to over 40% in both matchups - this is "slightly ahead," not "generational gap," and Tencent's own wording stays modest. What makes the methodology notable is the subject matter: expert blind ratings on real engineering tasks track the "actually gets work done" feel better than academic benchmarks do. It remains a Tencent self-evaluation, though - cross-validation will have to wait for the community to run it.

The README's Known Limitations section is refreshingly candid: it admits this is an early version of Hy4, with known issues including "spending longer than necessary reasoning through complex tasks" and "a tendency to over-verify its own work," and promises rapid iteration. Publishing your model's shortcomings on launch day sets a new bar for the open-source flagship race.

4. Recursive Self-Improvement: The Model Participated in Its Own Training

This is the real hook of the release. From the official press release: Hy4 preview participated "for the first time in the automated optimization of training methods, data strategies, evaluation frameworks, and low-level operators. The model proposed approaches, ran experiments, and iterated based on the results, with the resulting code, logs, and feedback feeding into subsequent rounds of exploration. This established an early-stage recursive self-improvement loop."

Unpacked, the model plays three roles in the loop - proposer, experiment runner, and result evaluator: it suggests improvements to the training pipeline it is itself being built with, runs the experiments, and iterates on the data, with the outputs feeding the next round. Tencent's disclosure here stays qualitative ("early-stage"); there is no quantified breakdown of how much these automated optimizations contributed to the final model's capability. Keep that boundary in view.

But one quantified data point is worth the record: Hy4 preview also autonomously analyzed bottlenecks in its own inference system and ran multiple rounds of optimization around operator fusion and communication, lifting end-to-end throughput 31.8% versus baseline, with consistent gains across different context lengths and concurrency levels. In other words, "the model optimizing its own inference infrastructure" has moved from vision to a deliverable with a number attached. The 2026 agent narrative has been stuck at "AI writes code, humans review" - this round moves it half a step forward: AI-written optimization code serving the model's own runtime.

5. Pricing and Access: $0.834 Input, Two Weeks Free on WorkBuddy/CodeBuddy

API pricing (official figures, per million tokens): $0.834 input, $2.501 output, $0.042 cache hit. The cache-hit price is 1/20th of the input price - friendly territory for agent-style, long-context workflows.

Item (per million tokens)Price
Input$0.834
Output$2.501
Cache hit$0.042

Access channels: WorkBuddy, CodeBuddy, Yuanbao, ima, and other Tencent products offer the model directly; the API runs through Tencent Cloud TokenHub and OpenRouter. From launch, Hy4 preview is free on WorkBuddy and CodeBuddy for two weeks, and free access to Hy3 on both platforms has been extended until September 30. For individual developers, the cheapest validation path is a round of real tasks inside CodeBuddy during the free window; if you want to self-host, jump straight to the companion SOP.

6. The Takeaway: The Open-Source Flagship Race Enters Its Self-Improving Half

Put Hy4 preview back into the open-source flagship lineup - GLM-5.3, Kimi K3, the DeepSeek V4 series - and the contest is no longer just leaderboard scores. It is a composite exam of parameter scale, deployment threshold, API pricing, and real performance on engineering tasks. Tencent's play this time: 770B fully open weights to show good faith, 49B activation plus MTP to control inference cost, expert blind evals to certify engineering capability, and a free window to lower the trial barrier.

The "participated in its own training" narrative matters more in the long run than any score. It signals that the flagship-model development process itself is starting to be rewritten by the model: evaluation frameworks, data strategies, and operators - stages once owned entirely by engineers - now carry the model's own contributions. Whether the loop is real and how much it delivers, time will tell; but the direction is already in the official press release: "The next batch of models in the Hy4 series is expected to roll out soon." Models building models will likely be the headline of that next batch.

One-line closer: when 770B weights ship fully open, 49B activation keeps costs in check, expert blind evals certify a slight lead - and the model helped train itself - the open-source flagship race has formally moved from "who scores higher" to "who gets the self-improvement loop running first."


References

This article is based on the official press release, GitHub README, and public reporting (as of 2026-08-29). Prices and star counts are snapshots; the official pages prevail. Not investment advice.

This article is AI-assisted and human-edited. Last updated: 2026-08-29

Related

Frontline Hotspot

280B Total, 14B Active, Three Regions Shut Out

On November 5, 2026 Baidu released ERNIE 5.5 Flash Base and published its weights: 280B total parameters with 14B activated per token, pretrained from scratch on 33.5T tokens, 128K training context and 1M for fine-tuning and inference, 43 layers with 37 heavily compressed attention layers plus 6 MLA-DSA layers, 512 routed experts with top-10 selection and one shared expert, and FP8 plus BF16 checkpoints on Hugging Face and ModelScope. This is an end-of-pretraining Base checkpoint with no mid- or post-training, not a chat model; a post-trained version is promised without a date. The custom ERNIE 5.5 Flash Base Model License Agreement v1.0 excludes the European Union, the United Kingdom and South Korea from its territory, allows free non-commercial use, requires a prominent Powered by ERNIE notice and a one-time notice for commercial use up to 50 million dollars in related annual revenue, requires separate prior written authorization above that threshold, and bans using outputs to improve unrelated AI systems.

Nov 6, 20269 min read
Frontline Hotspot

Kimi K3.1 Pro Weights Are Out: 2.8T Parameters, 1.56TB

At midnight Beijing time on November 5, 2026, Moonshot AI published the Kimi K3.1 Pro weights on Hugging Face, honoring the promise made when the model launched on October 23. The model has 2.8 trillion total parameters with about 104B activated per token, a 1048576-token context, 69 KDA plus 24 Gated DSA attention layers, a 401M vision encoder, and MXFP4 weights with MXFP8 activations under quantization-aware training, totaling roughly 1.56TB. The vendor-reported Terminal-Bench 4.0 score is 65.7; on the third-party Agent Arena of real user sessions it ranks first among open models with a net improvement rate of 5.29 percent, ahead of MiMo-V2.6-Pro at 2.68 percent (preliminary). The Kimi K3.1 Pro License requires a separate agreement for Model-as-a-Service businesses exceeding 20 million dollars in trailing 12-month revenue and prominent naming for products above 100 million monthly users or 20 million dollars monthly revenue.

Nov 5, 20269 min read
Frontline Hotspot

iFlytek Spark X2.6: A 293B Flagship, but Only a 4B Is Open

iFlytek launched Spark X2.6 at its global 1024 Developer Festival in Beijing on October 22, 2026: a 293B-parameter MoE that activates 30B per token, a 1M-token context window, an end-cloud integrated design, trained entirely on domestic compute and served from a 10,000-card domestic inference cluster. The efficiency ledger, all vendor-reported: training 2000B tokens of 256K-length data on the Feixing-2 platform fell from 4.5 months to 1.2 months (a 4x gain), and the first fully domestic four-layer KV Cache architecture lifted hit rate from 63 to 95.7 percent with 4.5x inference efficiency. Self-reported scores put SWE Pro at 77.3 and SWE Multilingual at 95.0, both first, with SWE Verified at 95.4 behind only DeepSeek-V4.1-Flash 1017's 95.6 - yet there is no complete score table, no technical report, and no third-party replication. The key correction: the viral headline 'iFlytek open-sources X2.6' misreads the release. Only the on-device Spark-X2.6-4B is open under Apache 2.0 (Hugging Face under the XHToken namespace, ModelScope mirror, Ollama as SparkLLM), while the 293B flagship stays API-only. The 4B self-reports six gains including BFCL-V4 71.8 and SWE Verified 62.8. Launch pricing is 0.9 / 0.08 / 3.6 yuan per million tokens. The narrative layer predicts 60 percent local, 30 percent hybrid, 10 percent cloud workloads with routing accuracy up from 81.2 to 96.1 percent; the chairperson's 'only model trained on fully domestic compute' is a unilateral claim with no chip model disclosed. Unpublished: chip models, cluster size, total FLOPs, technical report, flagship weights.

Oct 23, 20268 min read