Tencent Hy4 preview: 770B parameters, 49B active and a 1M context. Benchmarks, pricing and comparison with Kimi K3, GLM-5.3 and Claude Opus 5

On August 28, 2026 Tencent released and open-sourced Hy4 preview, a new flagship Mixture-of-Experts model in the Tencent Hy family, formerly better known internationally as Hunyuan.^[1][6]

Its specification is large even by 2026 standards:

770B total parameters
49B active parameters per token
78 layers
256 routed experts + 1 shared expert
top-8 routed experts active per token
1M-token context
Apache License 2.0

^[2]

Tencent published both the full model and Hy4 preview-FP8. The full Hugging Face repository is about 1.56 TB.^[2][3]

The company positions Hy4 preview as a model for real work rather than simple chat: software engineering, office workflows, financial analysis, game development, scientific research and tool-using agents.^[1][2]

Three details matter most.

First, the weights are released under Apache 2.0, so the model can be self-hosted and used commercially subject to that license.^[2][4]

Second, Hy4 posts strong vendor-reported coding and agent scores, including 82.9% on SWE-bench Multilingual, 85.4% on Terminal-Bench 2.1 and 83.7% on MCP-Atlas.^[2][9]

Third, Tencent says Hy4 preview participated in optimization of its own development and inference pipeline. The model reportedly proposed changes, ran experiments and iterated on results, while another optimization loop improved Tencent's end-to-end inference throughput by 31.8% versus its internal baseline.^[1]

That sounds like “AI improving itself”, but the careful wording matters: Tencent describes an early engineering-oriented recursive self-improvement loop, not a fully autonomous model training future generations without human oversight.

Information status: August 31, 2026.

TL;DR

Question	Verified answer
Model	Tencent Hy4 preview
Release	August 28, 2026
Type	Mixture-of-Experts
Total parameters	770B
Active parameters	49B per token
Layers	78
Routed experts	256
Shared experts	1
Active routed experts	top-8
Context	1M in model card, 1,048,576 on OpenRouter
Max output on OpenRouter	64K
Attention	Gated DeepSeek Sparse Attention
Sparse attention optimization	IndexCache
MTP	1 native layer, 10B total / 0.7B active
License	Apache 2.0
Full weights	about 1.56 TB on Hugging Face
Quantized variant	Hy4 preview-FP8
Self-hosting	vLLM and SGLang
Official recipe	tensor parallel size 8
API input	$0.834 / 1M tokens
API output	$2.501 / 1M tokens
Cache hit	$0.042 / 1M tokens
SWE-bench Multilingual	82.9%, Tencent-reported
SWE-bench Pro	65.7%, Tencent-reported
Terminal-Bench 2.1	85.4%, Tencent-reported
GPQA Diamond	92.3%, Tencent-reported
MCP-Atlas	83.7%, Tencent-reported
Status	Preview
Known limitations	over-long reasoning and over-verification
Main caveat	no broad independent replication of the launch benchmark table yet

What exactly did Tencent release?

Tencent made available:

Hy4 preview
Hy4 preview-FP8

on Hugging Face, ModelScope, GitCode and CNB.^[2]

The model is also accessible through Tencent products such as WorkBuddy and CodeBuddy, while API access is available through Tencent Cloud TokenHub and OpenRouter.^[1][7]

Reuters independently confirmed the August 28 launch and highlighted software engineering, research and financial analysis as the main stated use cases.^[6]

770B does not mean 770B active for every token

Hy4 preview uses a Mixture-of-Experts architecture.

The backbone contains:

770B total parameters

but roughly:

49B parameters

are activated for a given token.^[2]

That is about 6.4% of the backbone.

MoE lets the model hold a very large pool of parameters without executing all of them for every token.

It does not make the model lightweight. The full weights still have to be stored and distributed across inference hardware.

How are the experts arranged?

The backbone contains 78 layers.^[2]

The first layer uses a standard dense FFN.

The remaining 77 use MoE, each with:

256 routed experts
1 shared expert

For every token, the router activates:

top-8 routed experts
+
the shared expert

^[2]

That explains much of the gap between total and active parameters.

Hy4 vs Hy3: a major scaling jump

Hy3 used:

295B total
21B active
256K context

^[5][12]

Hy4 preview uses:

770B total
49B active
1M context

^[2]

That is approximately:

• 2.61× more total parameters,
• 2.33× more active parameters,
• at least 3.9× longer stated context.

Tencent says it scaled model size, context length and training data together, while also expanding post-training.^[2]

The 1M-token context

The model card lists:

Context Length: 1M

^[2]

OpenRouter exposes the model with:

1,048,576 token context
64,000 max completion tokens

^[7]

That context size is attractive for:

• large codebases,
• long agent sessions,
• multi-document workflows,
• research,
• repo-level refactoring,
• large logs and data.

But advertised context length is not the same as effective context quality. Retrieval, latency and memory behavior still need task-specific testing.

Gated DeepSeek Sparse Attention

Hy4 preview uses Gated DeepSeek Sparse Attention, inspired by DeepSeek's sparse-attention design.^[2]

Sparse attention attempts to reduce the cost of very long sequences by selecting relevant prior positions instead of evaluating full quadratic attention everywhere.

Tencent combines this with another optimization:

IndexCache

What does IndexCache do?

IndexCache reuses sparse-attention indices across layers.^[8]

A conventional sparse-attention indexer can select top-k relevant positions separately at many layers.

IndexCache exploits similarity between these selections so some layers reuse the indices calculated elsewhere.

The IndexCache paper reports reduced indexer work and prefill/decode speedups on the tested DSA models with small quality impact.^[8]

Those numbers are results for IndexCache research, not direct Hy4 benchmark results.

Native MTP layer

In addition to the backbone, Hy4 includes one native Multi-Token Prediction layer.^[2]

Tencent specifies:

10B total parameters
0.7B active parameters

for the MTP layer.

It is used for speculative decoding.

Tencent's official vLLM and SGLang recipes explicitly enable the MTP path.^[2]

The weights are genuinely public

This is not merely an API marketed as “open”.

Hugging Face shows:

tencent/Hy4-preview
1.56 TB
131 safetensors shards
Apache-2.0

^[3]

Tencent also publishes:

tencent/Hy4-preview-FP8

^[11]

This enables self-hosting, weight analysis, custom inference stacks, fine-tuning, quantization experiments and deployment without sending prompts to a third-party API.

Apache 2.0 matters more than the marketing label

The model card and official repository identify:

Apache License 2.0

^[2][4]

It is a permissive license that allows use, modification and redistribution, including commercial use, subject to its terms.

For enterprises, that is typically simpler than the custom community licenses used by some other open-weight models.

Official self-hosting: vLLM and SGLang

Tencent recommends:

vLLM
SGLang

for production serving.^[2]

The official vLLM recipe uses:

vllm/vllm-openai:hy4-preview

with the FP8 model and:

--tensor-parallel-size 8

^[2]

The SGLang recipe likewise uses --tp-size 8.

This does not mean “any eight GPUs are enough”. Memory requirements depend on precision, KV cache, context, concurrency and hardware type.

OpenAI-compatible API

Once served through vLLM or SGLang, Hy4 can be called through an OpenAI-compatible endpoint.^[2]

Example:

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8000/v1",
    api_key="EMPTY",
)

response = client.chat.completions.create(
    model="hy4-preview",
    messages=[
        {"role": "user", "content": "Review this repository architecture."}
    ],
    temperature=0.9,
    top_p=1.0,
)

Tencent recommends temperature=0.9 and top_p=1.0.^[2]

Reasoning can be reduced

The model defaults to a high reasoning mode for difficult tasks.^[2]

For more direct responses, the model card shows:

extra_body={
    "chat_template_kwargs": {
        "reasoning_effort": "no_think"
    }
}

That matters because one of Tencent's explicitly acknowledged preview issues is excessive reasoning duration.

Official preview limitations

Tencent lists two notable issues:^[2][6]

1. the model can spend longer than necessary reasoning through complex tasks,
2. it can over-verify its own work.

In agent workflows, extra verification may mean more tool calls, higher latency and more tokens.

Hy4 preview should therefore be treated as a powerful but still early release.

Benchmarks: the most important caveat first

Tencent published a benchmark appendix as part of the model card.^[2]

As of August 31, there is not yet broad independent, configuration-matched replication of the whole table.

We therefore label all benchmark values below as:

Tencent-reported

not:

independently verified

This matters particularly for agent benchmarks where tools, harness, token budget, reasoning settings and environment version can materially move results.

Selected Hy4 preview benchmark scores

A transcription of Tencent's benchmark appendix reports:^[2][9][10]

Benchmark	Hy4 preview
GPQA Diamond	92.3%
Terminal-Bench 2.1	85.4%
MCP-Atlas	83.7%
SWE-bench Multilingual	82.9%
MathArena Apex 2025	74.2%
SWE-bench Pro	65.7%
DeepSWE	64.3%
HLE, High + Tools	55.4%
HLE, High without tools	43.4%

All values above are vendor-reported.

SWE-bench Multilingual: 82.9%

Tencent reports:

Hy4 preview: 82.9%
Hy3: 75.8%

^[9]

That is a 7.1 percentage-point increase over Hy3 in the published configuration.

The benchmark is particularly relevant for software engineering across multiple languages.

It still should not be compared to arbitrary “SWE-bench” numbers without matching benchmark variant and harness.

SWE-bench Pro: 65.7%

Tencent reports:

Hy4 preview: 65.7%
Hy3: 57.9%

^[9][10]

That is a 7.8-point increase in Tencent's table.

It is one signal that Hy4 is a real software-engineering capability upgrade over Hy3, not just a parameter-count increase.

DeepSWE: the biggest generational jump

The benchmark appendix transcription lists:

Hy3: 28.0%
Hy4 preview: 64.3%

^[9]

That is an enormous generation-over-generation gap.

Precisely because it is so large, independent replication matters: scoring can be affected not only by the model but also the scaffold, tool use and evaluation configuration.

Terminal-Bench 2.1: 85.4%

Tencent reports:

Hy4 preview: 85.4%

on Terminal-Bench 2.1.^[9][10]

That is a strong result for a model aimed at coding agents.

However, Tencent's own appendix also shows several competitors at similarly high or higher values, so it would be incorrect to claim Hy4 universally “wins Terminal-Bench”.

GPQA Diamond: 92.3%

The vendor-reported result is:

92.3%

^[9][10]

This supports Tencent's positioning around scientific reasoning.

The company specifically names AI research, molecular dynamics, condensed-matter physics and mathematics as areas of improvement.^[1]

A benchmark score does not prove reliability in real scientific research.

MCP-Atlas and tool-using agents

Tencent reports:

MCP-Atlas: 83.7%

^[9][10]

That matters because Hy4 is intended for agent workflows and tool calling, not only text generation.

OpenRouter's integration supports tools, tool_choice and structured outputs via JSON schema.^[7]

Hy4 vs Kimi K3 and GLM-5.3: Tencent's blind test

Tencent also conducted an internal blind side-by-side evaluation.^[1][2]

Setup:

163 Tencent experts
203 engineering tasks
0–4 scoring scale

Average scores:

Hy4 preview: 2.99
Kimi K3: 2.94
GLM-5.3: 2.92

^[1]

Against Kimi K3 Tencent reports:

51.2% wins
7.9% ties
40.9% losses

Against GLM-5.3:

46.8% wins
12.8% ties
40.4% losses

^[2]

That is a narrow lead in Tencent's internal workload, not independent proof that Hy4 is universally superior.

What about Claude Opus 5 and GPT-5.6 Sol?

Tencent includes closed frontier models in its benchmark appendix.^[9]

Hy4 is close on some rows and behind on others.

For example, the transcription of Tencent's SWE-bench Multilingual row lists:

Hy4 preview: 82.9%
Claude Opus 5: 89.5 / 85.8%
GPT-5.6 Sol: 74.1%

and Terminal-Bench 2.1:

Hy4 preview: 85.4%
Claude Opus 5: 86.7 / 85.4%
GPT-5.6 Sol: 88.8 / 88.3%

^[9]

The double numbers represent multiple settings/variants shown in Tencent's chart. They should not be flattened into one universal ranking.

The safe conclusion is that Hy4 is competitive across many tasks, not that it is categorically the world's best model.

API pricing

Tencent lists:^[1]

Input:      $0.834 / 1M tokens
Output:     $2.501 / 1M tokens
Cache hit:  $0.042 / 1M tokens

OpenRouter shows the same input/output and cache-read pricing.^[7][13]

That is aggressive for a model of this scale.

But token price is not task price. A model that reasons longer or makes more tool calls can cost more per completed task despite a lower unit rate.

Self-hosting vs API

API is attractive when:

• you need fast adoption,
• you do not want to run a large GPU cluster,
• demand is variable,
• usage-based cost is acceptable.

Self-hosting is attractive when:

• data residency matters,
• you need inference control,
• you already operate GPU infrastructure,
• workload is large and stable,
• you want to control quantization and serving.

At roughly 1.56 TB for the full repository, this is not a typical single-workstation model.^[3]

Did Hy4 really “optimize itself”?

Tencent says the model participated in automated optimization of:^[1]

• training methods,
• data strategies,
• evaluation frameworks,
• low-level operators.

It reportedly proposed approaches, ran experiments and iterated using the results, with code, logs and feedback entering later rounds.

Tencent calls this an:

early-stage recursive self-improvement loop

The clickbait version:

Hy4 trained itself

is not supported.

The source describes automated assistance and iterative engineering inside an R&D process, not a fully autonomous self-replicating training system.

What does the +31.8% throughput claim actually mean?

Tencent says Hy4 analyzed bottlenecks in its inference system and iterated on operator fusion and communication optimization.^[1]

Tencent reports:

end-to-end throughput
+31.8%
vs internal baseline

This does not mean:

• 31.8% faster than Kimi,
• 31.8% lower latency than Claude,
• 31.8% cheaper than competing APIs.

It is a gain against Tencent's internal inference baseline.

Without that qualifier the number is easy to misread.

Biggest strengths

Permissive license

Apache 2.0.

Very long context

1M tokens.

High intelligence density

49B active from 770B total.

Strong vendor-reported coding/agent scores

Especially SWE-bench Multilingual, Terminal-Bench 2.1 and MCP-Atlas.

Official serving recipes

vLLM and SGLang.

FP8 variant

Lower memory cost than full BF16 weights.

Competitive API price

Below many closed frontier offerings.

Main risks and unknowns

Preview status

It is still an early release.

Benchmark independence

The strongest launch numbers are Tencent-reported.

Over-reasoning

Officially acknowledged.

Over-verification

Can increase agent latency and cost.

Infrastructure

49B active does not eliminate the need to store a huge model.

1M context

Recall, latency, KV-cache use and quality must be tested separately.

Data governance

Hosted API terms must be checked provider by provider. Open weights do not automatically make a hosted endpoint compliant with an organization's privacy requirements.

What should a fair enterprise benchmark look like?

Do not merely copy the vendor chart.

Compare:

Hy4 preview
Kimi K3
GLM-5.3
DeepSeek V4
Claude Opus 5
GPT-5.6 Sol

on identical internal tasks.

Measure:

• task completion rate,
• test pass rate,
• final diff correctness,
• tool calls,
• retries,
• input/output tokens,
• latency,
• cost per completed task,
• regressions,
• out-of-scope changes,
• long-context quality.

The most useful metric is often:

cost per successful task

rather than:

cost per million tokens

Deployment checklist

Model quality

• Test on internal tasks.
• Separate Tencent benchmark claims from independent results.
• Compare high reasoning and no_think.
• Measure over-verification.
• Test hallucinations in your domain.
• Validate tool calling.
• Validate structured outputs.
• Test long agent sessions.

Long context

• Test 32K, 128K, 256K and larger contexts.
• Measure retrieval accuracy.
• Measure time-to-first-token.
• Measure KV-cache footprint.
• Do not equate 1M advertised context with 1M useful memory.
• Test repo-level coding.
• Test cross-document reasoning.
• Test prompt injection in long context.

Self-hosting

• Verify full GPU requirements.
• Start with official FP8.
• Test vLLM.
• Test SGLang.
• Measure throughput at real concurrency.
• Measure P50/P95/P99 latency.
• Plan storage for weights.
• Review Apache 2.0 obligations with legal counsel.

API

• Verify current provider pricing.
• Check cache semantics and TTL.
• Check data retention.
• Check processing region.
• Set cost limits.
• Monitor usage per agent.
• Maintain a fallback model.
• Measure cost per successful task.

Is it worth upgrading from Hy3?

For teams already using Hy3, Hy4 preview is an obvious POC candidate.

On paper:

295B → 770B total
21B → 49B active
256K → 1M context

plus large vendor-reported improvements on coding and agent evaluations.^[2][5][9]

But an upgrade should not be automatic.

Hy4:

• may reason longer,
• requires more self-hosting infrastructure,
• is still preview,
• can have different latency and cost behavior.

Production migration should be driven by real workloads, not only a benchmark image.

POLPROG verdict

Hy4 preview is one of the most important open-weight releases of late August 2026.

Not because 770B is an impressive headline by itself.

The compelling combination is:

770B total
49B active
1M context
Apache 2.0
FP8
vLLM + SGLang
aggressive API pricing
strong coding/agent profile

It shows how quickly open models are moving from “cheaper alternative” to serious candidates for enterprise workloads.

Three evidence levels should remain separate.

Technical facts

Parameters, architecture, license, weights, context, deployment and pricing are well documented by Tencent, Hugging Face and providers.^[1][2][3][7]

Benchmarks

They are impressive but remain largely Tencent-reported at launch time.^[2][9][10]

“Self-improvement”

It is a real part of Tencent's described R&D workflow, but should not be presented as fully autonomous self-training.^[1]

The most reasonable conclusion on August 31, 2026 is:

Hy4 preview belongs on the shortlist for teams building coding agents, research agents, long-context workflows and self-hosted AI infrastructure.

There is not enough evidence yet to say:

Hy4 is the best model in the world.

There is enough evidence to say:

Tencent has pushed Hy4 into the small group of open-weight models that cannot be ignored when designing a production AI stack.