Tencent Hy4 Preview Benchmarks: 770B Open Model Takes on Kimi K3

Tencent has spent 2026 turning its Hunyuan lab into a serious open-weight contender. On August 28, 2026, the company released Hy4 preview, a 770-billion-parameter mixture-of-experts model published under the Apache 2.0 license on the same day it launched, with 49 billion parameters active per token and a native 1-million-token context window.

Tencent’s benchmark appendix lines Hy4 preview up against the current frontier field: DeepSeek V4 Pro 0813, Qwen 3.8 Max, GLM 5.3, Kimi K3, GPT-5.6 Sol, and Claude Opus 5. The numbers tell a nuanced story, one of genuine open-weight gains alongside claims that no independent lab has reproduced yet. Here is the full Tencent Hy4 preview breakdown, plus what the benchmarks do and do not prove.

What Is Tencent Hy4 Preview?

Hy4 preview is the next-generation flagship of Tencent’s Hunyuan (“Hy”) model line, developed by the in-house Tencent Hy Team in Shenzhen. It is a sparse mixture-of-experts transformer: 78 layers, with one dense feed-forward layer followed by 77 MoE layers, each routing across 256 experts plus one shared expert with only the top 8 activated per token.

Attention uses a gated variant of DeepSeek-style sparse attention with an index cache, and the model ships with a built-in multi-token prediction layer for faster speculative decoding.

A few specifications define what the model is for:

• Total parameters: 770 billion, with 49 billion active per token
• Context window: 1,048,576 tokens (1 million), four times its predecessor
• Modality: text in, text out only, no vision
• License: Apache 2.0, with BF16 (~1.56TB) and FP8 weights on Hugging Face, ModelScope, GitCode, and CNB
• Serving: day-zero vLLM and SGLang support, plus hosted API access
• Reasoning: on by default, with a no-think opt-out route
• Architecture details: Gated DeepSeek Sparse Attention with a cross-layer index cache, and four parallel residual streams for training stability at depth
• Deployment weight: roughly 1.56TB of BF16 weights, or 753GB-plus in the FP8 quantized build
• Ecosystem: Tencent’s CodeBuddy and WorkBuddy coding assistants run on the same Hunyuan line
• Evidence status: all benchmark figures remain vendor-reported pending independent reproduction

Tencent positions Hy4 preview as a “preview” release, with the full Hy4 model expected later. How should readers interpret “preview” in practice? It signals that both capability and pricing remain in flux, and that today’s benchmark snapshot may not represent the model developers will actually deploy against. That framing matters for two reasons. First, benchmark scores recorded on preview weights may shift again when the finalized model ships. Second, releasing preview weights under a permissive license is a deliberate strategy: it gets developers building on the Hy4 generation now, months before a finalized release, mirroring how Alibaba and Zhipu have treated recent community builds.

Hy3 vs Hy4 Preview: What Changed

Hy3 arrived on July 6, 2026 with 295 billion total parameters, 21 billion active, and a 256K context window. Hy4 preview roughly doubles active capacity, multiplies total parameters by about 2.6, and quadruples the context window, all within seven weeks. The schedule is the story: this is not a lab resting on a flagship for a year, but shipping a measurable leap every few weeks.

The capability gap is wider than the spec gap. On Tencent’s own table, the largest generational gains land on long-horizon agentic tasks: DeepSWE jumps from 28.0 to 64.3, SWE-Marathon from 5.0 to 31.9, and MathArena Apex 2025 from 38.7 to 74.2. According to the Tencent Hy Team’s model card, the leap reflects both a stronger pre-training recipe and a substantially larger post-training run.

The table below isolates that generational leap. It compares Hy3 against Hy4 preview on every benchmark Tencent published, with no rival models mixed in, so the scale of the jump is easy to read row by row.

Hy4 Preview vs the Frontier: The Full Benchmark Breakdown

Tencent published more than 40 benchmark rows spanning agentic coding, agentic search, working-agent tasks, STEM agents, and pure reasoning. The tables below split that appendix by task family so each one fits the page, and they drop the Hy3 column, since the generational comparison already has its own table above. One provenance note before the tables: values marked with an asterisk were reproduced by Tencent on its own evaluation setup rather than taken from the benchmark developer, which makes them vendor-controlled comparisons, not neutral third-party results.

“Stronger pre-training and a substantially larger post-training run compound into another step change in capability: the largest generation-over-generation gain we’ve measured.”

Tencent Hy Team, Hy4 preview model card

Agentic Coding

Agentic Search

Working Agent

STEM and Reasoning

A row-by-row audit of the tables shows where the “beats the frontier” framing holds and where it breaks down:

• Where Hy4 preview leads: it posts the top open-model score on the SWE Atlas coding trio (64.0 Q&A, 57.8 test writing, 53.3 refactoring), JobBench (61.7), BioMysteryBench (71.3), and internal search benchmarks, and it edges GPT-5.6 Sol on SWE-bench Pro (65.7 vs 64.6) and SWE Atlas Refactoring (53.3 vs 52.4) in Tencent’s re-runs.
• Where it roughly ties: Terminal-Bench 2.1 at 85.4 sits in a statistical dead heat with the entire frontier cluster (GLM 5.3 at 88.2, Kimi K3 at 88.3, GPT-5.6 Sol at 88.8 on reported figures), and GPQA Diamond at 92.3 is within a point of every rival.
• Where it trails: SWE-Marathon (31.9 vs 42.5 for GLM 5.3 and Kimi K3, 50.0 for Claude Opus 5), CritPt (16.9 vs 23.4 for Kimi K3 and 32.3 for GPT-5.6 Sol), MathArena Apex 2025 (74.2 vs 90.0+ for the closed leaders), AutomationBench (32.1 vs 48.2 for GLM 5.3), WorkspaceBench (60.2 vs 68.2+), and GDPval-AA Elo (1678 vs 1763 for GLM 5.3 and 1831 for Claude Opus 5).
• Where the closed frontier stays dominant: Claude Opus 5 holds large margins on SWE-bench Pro (79.2 vs 65.7), NL2Repo-Bench (75.3 vs 58.9), ProgramBench (39.5 vs 17.5), and BrokenArXiv (77.7 vs 54.6).
• Where agentic search is a dead heat: WideSearch, OneMillionBench, and DRACO all fall within a few points across the entire field, making the agentic search tier the least differentiated category in Tencent’s appendix.
• Where internal benchmarks favor the home team: Hy-SWE Max Verified, Hy-CompanyBench V2, and Hy-Backend 2.0 are Tencent-designed tasks; GLM 5.3 and GPT-5.6 Sol outscore Hy4 preview on all three, a result worth weighing when interpreting home-field advantage.

For context on the comparison models, Intelligent Living has covered the Qwen3.8 Max update in depth, along with the DeepSeek V4.1 Flash vs GLM 5.3 Flash matchup that defines the open-model price war Hy4 preview is entering.

Pricing and Availability: Where to Try Hy4 Preview

Like much of the Chinese AI field, Tencent is competing on price as much as capability, a pattern visible everywhere from budget releases like Qwen3.8-Flash to full flagships. Hy4 preview API access runs through Tencent Cloud TokenHub (priced in yuan at 6 per million input tokens, 18 per million output, 0.3 per million cache hits) and through OpenRouter, where the listed rates convert to roughly:

Access options break down as follows:

• Hosted API: Tencent Cloud TokenHub endpoints (China and US regions) and the Hy4 preview listing on OpenRouter, at $0.834 per million input tokens, $2.501 per million output, and $0.042 per million cache reads
• Consumer products: bundled inside Tencent’s WorkBuddy, CodeBuddy, Yuanbao, and ima apps, with two weeks of free access at launch
• Self-hosting: BF16 and FP8 weights on Hugging Face, ModelScope, GitCode, and CNB, with day-zero support merged into the vLLM serving library plus an official SGLang image
• Legacy access: Tencent extended free use of Hy3 through September 30, 2026
• Sampling defaults: official guidance is temperature 0.9 and top_p 1.0, with reasoning on by default and a documented opt-out route that differs between self-hosted and TokenHub endpoints
• Tooling: the release notes list Cline, OpenCode Go, and Vercel AI Gateway among tools supporting the model from day one

Self-hosting is where the open license meets hardware reality. The BF16 checkpoint spans roughly 1.56TB, and even the FP8 build requires more than 753GB of weight storage plus an 8-GPU node to serve, which puts local deployment firmly in data-center territory. Teams without that footprint will route through hosted API endpoints instead, which is where the pricing table above becomes the deciding factor.

The pricing pattern also fits a broader trend: Chinese labs consistently undercut Western frontier APIs while publishing open weights, a strategy we examined in our analysis of why Chinese AI models are so much cheaper than offerings from OpenAI and Anthropic.

The Verification Gap: What the Benchmarks Do Not Prove

Here is the part most coverage skips. Every number in Tencent’s comparison table is vendor-reported, run on Tencent’s own evaluation setup, and unreproduced by any independent lab. Asterisked competitor scores are Tencent’s own re-runs of rival models, meaning the comparison is only as trustworthy as Tencent’s harness.

The headline blind evaluation, which scored Hy4 preview 2.99 out of 4.00 against GLM 5.3’s 2.92 and Kimi K3’s 2.94, used 163 internal Tencent experts rating outputs on 203 internal engineering tasks, and margins that narrow could plausibly flip on a differently sampled task set.

Sorting the claims by evidence quality gives a clearer picture. Three tiers matter, from independently checkable evidence down to numbers that exist only on Tencent’s own slide decks:

Why does this distinction matter in practice? Because infrastructure teams making deployment decisions this month cannot wait for peer review; they need to know which claims carry independent corroboration and which rest on vendor documentation alone.

• Independently corroborated: OpenRouter pricing listings, the merged vLLM pull request (#54160) enabling day-zero serving, the downloadable Apache 2.0 weights themselves, and a LMArena Code Arena WebDev placement of roughly #5 overall (about 1633 points, #3 among open models) that jumped from Hy3’s #31 overall spot
• Vendor-reported only: all 40-plus benchmark rows, the internal blind evaluation, and Tencent’s claim that Hy4 preview optimized parts of its own training and inference stack, boosting end-to-end throughput 31.8%
• Not yet rated: the Artificial Analysis Intelligence Index, the closest thing the industry has to a neutral cross-lab scoreboard, had no published Hy4 preview entry as of early September 2026, with an independent evaluation listed as forthcoming

Tencent is also candid about weaknesses in its own model card: Hy4 preview spends longer than necessary reasoning through complex tasks and tends to over-verify its own work, traits that inflate token costs exactly where agentic workloads get expensive. The HLE numbers reinforce the pattern, since Hy4 preview scores 55.4 with tools but 43.4 without, one of the widest tool-dependency gaps in the field.

That candor cuts both ways, because naming known flaws reads as transparency while simultaneously lowering expectations for what preview weights can deliver. There is a genuine tension in the release narrative worth naming: Tencent describes the model contributing to its own training and inference optimization, a recursive self-improvement claim, while simultaneously publishing benchmark tables that no external party has reproduced. The arena placement suggests the coding gains are real, but anyone choosing a model on Tencent’s benchmark table alone is reading a vendor’s scorecard, not an audit.

Why Hy4 Preview Matters for Open-Weight AI

Evidence quality aside, the release is significant for three reasons that have little to do with any single leaderboard position. Each speaks to how the open-weight ecosystem is evolving, rather than to where Hy4 preview ranks this month.

The license is genuinely permissive. Apache 2.0 ships without the revenue thresholds, monthly-active-user caps, or model-as-a-service carve-outs attached to several rival community licenses, making Hy4 preview one of the least-restricted open-weight flagships of the season. For builders, license fine print matters as much as benchmark scores, and the permissive terms directly enable the self-hosting ecosystem described earlier in this article.

Open models are openly iterating on each other. Hy4 preview’s attention mechanism adapts DeepSeek’s sparse attention design, and Tencent’s model card credits DeepSeek and GLM as inspirations. The Chinese open-weight ecosystem compounds: architectures published by one lab become the next lab’s foundation, which is precisely how open releases like Qwen 3.8’s 27B open model have ended up within striking distance of closed frontier systems.

The model reportedly helped build itself. Tencent states that Hy4 preview contributed to optimizing parts of its own training and inference systems, autonomously tuning operators and communication patterns to produce that 31.8% throughput gain. If verified, that kind of recursive self-improvement loop in an openly released model is a milestone worth watching independently of any leaderboard.

Frequently Asked Questions

Is Tencent Hy4 preview free to use?

API access is not free but is inexpensive: $0.834 per million input tokens and $2.501 per million output tokens via TokenHub or OpenRouter. One caveat on token budgets: because the model overthinks complex tasks by default, effective costs on agentic workloads can run higher than the headline per-token price suggests. Tencent offered two weeks of free access through WorkBuddy and CodeBuddy at launch, extended free Hy3 access through September 30, 2026, and self-hosting the open weights carries no license fee at all.

What license is Hy4 preview released under?

Apache 2.0, one of the most permissive open-source licenses available. Weights ship as BF16 and FP8 checkpoints on Hugging Face, ModelScope, GitCode, and CNB, with no revenue cap or user-count restriction.

Is Hy4 preview better than Kimi K3?

On Tencent’s own blind evaluation, yes, but barely: 2.99 out of 4.00 vs 2.94, with mixed head-to-head win rates. The published benchmark table shows a split picture, with Hy4 preview ahead on several agentic coding and search tasks but behind Kimi K3 on SWE-Marathon, CritPt, AutomationBench, and several working-agent benchmarks. Independent verification is pending.

Have independent labs verified the Hy4 preview benchmarks?

Not the benchmark table itself. All 40-plus rows are vendor-reported, and Artificial Analysis had not published an Intelligence Index score for the model as of early September 2026. The closest independent signal is a LMArena Code Arena WebDev placement of about #5 overall and #3 among open models.

What hardware do I need to run Hy4 preview locally?

An 8-GPU server node at minimum. The FP8 checkpoint alone exceeds 753GB of weights, and the BF16 build is roughly 1.56TB. Day-zero vLLM and SGLang images handle serving, with an Ascend NPU path also documented, but this is data-center hardware, not a workstation model, unlike lighter open-weight releases such as DeepSeek V3.2 built for local reasoning.

What are Hy4 preview’s known weaknesses?

Tencent’s own model card flags two: the model overthinks complex tasks, burning more tokens than necessary, and it tends to over-verify its own work, which compounds in agentic loops. It is also text-only, with no vision or audio support in this preview release.

Conclusion

On Tencent’s numbers, Hy4 preview is the strongest open-weight coding model China has produced, trading blows with Kimi K3, GLM 5.3, and DeepSeek V4 Pro while occasionally edging GPT-5.6 Sol, and it does so under a clean Apache 2.0 license at roughly a tenth of closed-frontier output pricing. The honest read is more measured: the capability claims remain vendor-reported, the biggest leads land on Tencent’s own internal benchmarks, and the model trails badly on long-horizon and elite reasoning tasks. The independent scoreboard is coming, from Artificial Analysis and from the open community now running the weights on their own hardware, where open models have already proven capable of rivaling far larger systems. When those results land, we will know whether Hy4 preview is genuinely frontier-class or merely frontier-priced.