Hy4 preview API

Hy4 preview is a new-generation Mixture-of-Experts (MoE) flagship model developed by the Tencent Hy team, released on August 28, 2026. It has 770B total parameters, of which 49B are activated per token. The backbone has 78 layers: the first uses a standard dense FFN and the remaining 77 use MoE layers, each with 256 routed experts and 1 shared expert. Every token activates the top 8 routed experts along with the shared expert. A native Multi-Token Prediction (MTP) layer with 10B total and 0.7B activated parameters is built in for speculative decoding.

On the architecture side, and inspired by DeepSeek and GLM, the attention module uses Gated DeepSeek Sparse Attention (Gated DSA) with IndexCache for cross-layer reuse of sparse indices. The residual pathway uses iHC (identity Hyper-Connections) to widen the flow of information between layers. Hy4 preview is released under the Apache License 2.0.

A New Flagship Generation

Tencent scaled Hy4 preview on three fronts: model size, context length and training data. Stronger pre-training and a substantially larger post-training run compound into what Tencent calls the largest generation-over-generation gain it has measured, enough to put Hy4 preview at the open-source frontier.

Built for Productivity

Tencent partnered with experts inside the company, such as software engineers, game developers, finance analysts and security experts. It built training data around the work they ship:

• Software engineering. Better at understanding, planning, debugging and checking long-horizon development tasks, with further gains in the visual taste and interaction quality of front-end work.
• Office and analysis. Takes messy context spread across many files and turns it into shareable documents, spreadsheets and presentations, handling data analysis, equations and financial models with greater precision.
• Game development. Turns a single prompt into a playable prototype and works fluently with game engines, so developers can keep refining complex projects over multiple turns.
• Scientific research. Stronger understanding, reasoning and problem solving on hard research questions, with progress across AI research, molecular dynamics, condensed matter physics and pure mathematics.

Tencent co-designs Hy4 preview with its own products such as CodeBuddy and WorkBuddy. In a blind side-by-side evaluation, 163 internal experts rated model outputs on 203 engineering tasks. Hy4 preview came out slightly ahead of GLM 5.3 (2.99 vs 2.92 average, 46.8% wins, 12.8% ties and 40.4% losses) and of Kimi K3 (2.99 vs 2.94, 51.2% wins, 7.9% ties and 40.9% losses).

Model Specifications

Backbone parameters only, excluding the MTP layer. All figures are from Tencent's official model card.

Property	Value
Architecture	Mixture-of-Experts (MoE)
Total Parameters	770B
Activated Parameters	49B
Layers	78
Hidden Size	6144
Attention Type	Gated DSA
Attention Heads	64
Query Compression Dimension	2048
Key-Value Compression Dimension	512
Indexer Heads / Head Dimension	32 / 128
Indexer top-k	2048
Residual Streams	4
Routed Experts	256
Shared Experts	1
Activated Routed Experts per Token	8
MoE Intermediate Size	2048
FFN Intermediate Size	18432
Context Length	1M
Vocabulary Size	120832

Benchmarks

Reproduced from the benchmark appendix in Tencent's official model card. Each model is evaluated at its highest available reasoning setting. Results marked * are from Tencent's own testing, and where a cell holds two values both are as published. "n/a" means no score is reported.

Agentic Coding

Benchmark	Hy3	Hy4 preview	DeepSeek V4 Pro 0813	Qwen 3.8 Max	GLM 5.3	Kimi K3	GPT 5.6 Sol	Claude Opus 5
SWE-bench Multilingual	75.8	82.9	77.3*	82.6*	81.3*	80.8*	74.1*	89.5/85.8*
SWE-bench Pro	57.9	65.7	60.3*	67.7/61.6*	64.6*	63.3*	64.6/60.5*	79.2/79.9*
DeepSWE	28.0	64.3	62.7/58.8*	56.6/55.6*	66.9/68.1*	67.5/74.0*	72.7/68.9*	68.8/74.7*
SWE Atlas: Codebase Q&A	30.8	64.0	53.4*	55.4*	55.8*	35.2*	58.1*	57.5*
SWE Atlas: Test Writing	35.9	57.8	45.6*	52.8*	49.6*	35.6*	49.6*	68.7*
SWE Atlas: Refactoring	32.9	53.3	48.6*	51.0*	51.9*	37.4*	52.4*	60.0*
SWE-Marathon	5.0	31.9	19.0*	31.0*	42.5/35.6*	42/44.4*	42.5/35.9*	50.0/48.0*
Terminal-Bench 2.1	70.8	85.4	87.9/80.3*	86.6/85.8*	88.2/88.3*	88.3/85.7*	88.8/88.3*	86.7/85.4*
NL2Repo-Bench	45.6	58.9	61.5/54.4*	55.9/58.0*	58.0/56.1*	58.0/58.3*	56.8*	75.3*
CyberGym	51.8	78.4	83.3/80.3*	78.5/78.5*	84.5/83.0*	80.0	83.6	n/a
ProgramBench	3.0	17.5	15.5*	17.5*	18.0*	24.5*	25.0*	39.5*
PostTrainBench V1.1	14.5	35.6	24.5*	n/a	33.2*	32.0*	36.2	35.0
Harbor-Index	15.6	39.6	36.9*	38.8*	42.5*	n/a	46.3*	56.9*
Hy-Backend 2.0 (Internal)	26.2	35.2	33.8*	34.9*	41.9*	37.9*	49.6*	40.3*
Hy-SWE Max Verified (Internal)	49.0	64.2	65.7*	65.2*	67.2*	65.6*	69.8*	70.1*
Hy-CompanyBench V2 (Internal)	29.8	62.4	64.4*	63.3*	64.5*	63.3*	70.6*	72.7*

Agentic Search

Benchmark	Hy3	Hy4 preview	DeepSeek V4 Pro 0813	Qwen 3.8 Max	GLM 5.3	Kimi K3	GPT 5.6 Sol	Claude Opus 5
WideSearch	81.9	83.9	81.8*	81.9/81.1*	83.2*	81.0*	86.3*	84.0*
OneMillionBench (with tools)	51.5	65.4	62.0*	63.1*	64.5*	63.5*	67.1*	68.1*
DRACO	65.2	77.2	77.3*	76.4*	78.1*	77.5*	77.7*	88.6/79.2*
Hy-LifeSearch (Internal)	38.9	49.2	46.9*	47.5*	49.2*	45.8*	63.4*	56.1*
Hy-BrowseComp-Pro2 (Internal)	55.0	56.1	46.5*	46.7*	48.4*	58.1*	56.4*	61.3*

Working Agent

Benchmark	Hy3	Hy4 preview	DeepSeek V4 Pro 0813	Qwen 3.8 Max	GLM 5.3	Kimi K3	GPT 5.6 Sol	Claude Opus 5
OfficeQA Pro	54.1	66.2	65.4*	65.4*	66.2*	65.4*	65.4*	66.9/66.9*
MCP-Atlas (public)	75.0	83.7	82.5*	81.9*	81.9*	84.2/82.8*	82.5*	85.7*
Toolathlon-Verified	56.2	74.1	74.1/70.1*	72.5/69.1*	73.0/73.8*	76.5/74.7*	73.2*	76.5*
APEX-Agents (pass@1)	24.4	37.1	32.4*	34.0*	38.1*	41.0/37.2*	39.9/37.9*	41.8*
SkillsBench (79, text-only)	55.3	62.9	65.0*	66.7*	63.3*	51.9*	62.5*	63.7*
JobBench	34.6	61.7	54.1*	53.0/52.2*	58.2*	54.3/54.7*	45.4/46.8*	68.0*
WorkspaceBench	58.2	60.2	65.4*	67.7/66.9*	68.2*	65.0*	65.3*	75.0*
Agents' Last Exam (ALE-CLI)	17.1	22.8	21.9*	25.4*	23.8*	23.2*	27.6*	25.1*
GDPval-AA V2 (Elo, official)	1213	1678	1580	1717	1763	1675	1711	1831
AutomationBench (v1.0.6)	16.1	32.1	30.4*	39.8/41.1*	48.2/49.4*	46.7/45.5*	45.8/39.9*	48.7*
BankerToolBench	68.8	78.6	73.1*	74.7*	77.8*	73.5*	79.0*	81.9*
E-Bench (Internal)	48.5	77.1	61.3*	66.8*	71.4*	73.8*	80.6*	77.8*
E-Bench-Code (Internal)	64.4	79.0	64.3*	67.1*	66.5*	77.6*	83.3*	82.7*
Hy-FinAgentBench (Internal)	69.5	79.7	78.5*	77.2*	80.4*	78.5*	83.0*	82.0*
Hy-FinmodelBench v2 (Internal)	28.6	57.0	51.3*	52.5*	57.8*	52.4*	65.3*	66.0*

STEM Agent

Benchmark	Hy3	Hy4 preview	DeepSeek V4 Pro 0813	Qwen 3.8 Max	GLM 5.3	Kimi K3	GPT 5.6 Sol	Claude Opus 5
BioMysteryBench	54.9	71.3	61.6*	58.9*	69.0*	61.3*	73.1*	72.1*
HLE (with tools, text-only)	51.9	55.4	60.0/55.8*	56.2/54.1*	62.5/54.3*	57.0*	60.2*	60.9*

Reasoning

Benchmark	Hy3	Hy4 preview	DeepSeek V4 Pro 0813	Qwen 3.8 Max	GLM 5.3	Kimi K3	GPT 5.6 Sol	Claude Opus 5
CritPt (official)	4.9	16.9	18.0	20.0	19.1	23.4	32.3	29.1
GPQA Diamond	90.9	92.3	92.8/91.7*	92.6/92.2*	91.7/91.4*	93.5/92.8*	94.1/94.7*	93.7/93.3*
HLE (no tools, text-only)	34.4	43.4	42.7/40.5*	43.6/41.5*	42.3	46.9/46.6*	49.5/49.6*	54.9/53.2*
SUPERChem	52.6	66.4	62.0*	61.9*	58.5*	66.9*	73.6*	76.7*
ArXivMath	51.7	66.6	62.1*	67.1*	n/a	60.8*	79.5*	71.5*
HorizonMath (pass@4)	3.5	8.8	4.42*	5.31*	n/a	7.08*	10.62*	5.3*
MathArena Apex 2025	38.7	74.2	66.3*	72.8*	n/a	68.4*	90.0*	91.4*
BrokenArXiv	26.7	54.6	43.1*	42.7*	n/a	56.3*	64.4*	77.7*

Benchmark methodology notes (as disclosed by Tencent)

• Certain Hy3 scores may differ from previously reported results following updates to the evaluation harness, judge model and anti-hacking mechanisms.
• SWE-bench Multilingual and SWE-bench Pro. Evaluated with the swe-agent scaffold.
• DeepSWE. Evaluated with mini-swe-agent. Each task runs in a sandbox with 8 CPU cores and 16 GB of memory.
• SWE-Marathon. Official settings, except that the agent timeout is set to twice the official value. To reduce variance each task is run repeatedly until eight valid scores from successful runs are obtained.
• Terminal-Bench 2.1. Evaluated with the Claude Code harness, allowing up to 500 turns and a 12-hour timeout per trial. Resources are capped at 16 CPUs and 32 GB of RAM.
• MCP-Atlas. Follows Scale's April 2026 methodology on the 500-task public set, keeping the 100 tool-call budget per task and updating the judge model from Gemini 2.5 Pro to Gemini 3.1 Pro Preview.
• SkillsBench. All models are evaluated with Claude Code on 79 self-contained tasks, excluding multimodal tasks, averaged over 3 runs.
• SWE Atlas. The rubric judge parsing failure is fixed and a network allowlist prevents hacking behaviors. All models are evaluated with Claude Code with a 256-turn budget.
• NL2Repo. All models are evaluated with Claude Code with a 1000-turn budget per task, with additional anti-hacking prompt constraints and tool-call monitoring.
• ProgramBench. All models are evaluated with Claude Code with a 2000-turn budget per task. DeepSeek-V4-Pro-0813 uses mini-swe-agent instead, because its thinking frequently got stuck in a repetitive loop.
• Hy-SWE Max Verified (Internal). An in-house benchmark of 300 tasks run with the Claude Code scaffold and a 200-turn budget in a sandbox with 16 CPU cores and 32 GB of memory, with web access disabled and image reading blocked. Results are averaged over 3 runs and scored pass or fail by hidden tests.
• Agents' Last Exam (CLI). The 105 CLI tasks are evaluated with the official protocol and the Claude Code harness. Each task runs for up to 12 hours in a sandbox with 8 CPU cores and 32 GB of memory and is scored by the official ALE evaluators.
• APEX-Agents. Evaluated with the official ReAct Toolbelt harness and a maximum of 250 steps per trial, with resources capped at 4 CPUs and 16 GB of RAM.
• BioMysteryBench. Models are evaluated with Claude Code under a per-task budget of 2 hours and 250 turns, except GPT-5.6 Luna and GPT-5.6 Sol, which use Codex. All 99 tasks of the initial release are judged by Kimi K3.
• OfficeQA Pro. Claude Code is the agent harness for all models except the GPT series, which uses Codex CLI.
• Toolathlon-Verified. Uses Tencent's internal agent scaffold with some MCP tools reimplemented, each task pinned to 2 CPU cores and 10 GB of memory and a 2-hour timeout. The 108-task set, official programmatic evaluators, 100-step budget, 3-run Pass@1 protocol and highest reasoning effort are unchanged.
• Agentic search. All models are evaluated with Tencent's internal harness.
• BankerToolBench. All models are evaluated with the OpenCode scaffold with Gemini 3 Flash as the judge.
• Reasoning benchmarks. Five Claude Opus 5 results use high-setting runs while the others use max-setting runs. The high-setting runs have truncation rates of 2.32% on HLE, 15.17% on ArXivMath, 28.76% on HorizonMath, 5.86% on MathArena Apex 2025 and 17.17% on BrokenArXiv.

Reasoning and Sampling

Hy4 preview reasons by default at Tencent's high setting, a deep chain-of-thought mode suited to math, coding and reasoning. It also offers a no-think mode for direct answers. On Run BiOS reasoning is on by default at high and reasoning_effort: "none" switches to direct answers. Tencent recommends temperature = 0.9 and top_p = 1.0.

Known Limitations

Tencent describes Hy4 preview as an early version of Hy4 with real headroom left in both pre-training and post-training. It ships with known issues, among them spending longer than necessary reasoning through complex tasks and a tendency to over-check its own work. Tencent says it will keep iterating quickly on these. The official release, Hy4, is available on Run BiOS as tencent/hy4.

On Run BiOS

Hy4 preview is available on Run BiOS as the serverless model tencent/hy4-preview, with a 1,000,000-token context window, up to 64,000 output tokens, text input, function calling, structured outputs, reasoning that can be switched off and automatic prompt caching.