[whatmodel.sh](../../README.md) / Hy4 preview (770B-A49B) # Hy4 preview (770B-A49B) Tencent · released 2026-08 · 770B params (49B active) · apache-2.0 **score 76.7** ±1.0 · #21 of 732 · from 15 benchmark results across 2 sources Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only. ## What it's good at _{● marks Hy4 preview (770B-A49B) among every model on that task or test, worst to best.}

task         vs the others        rank  counts  tests
agent tasks  ▁▁▃▆█▄▅▅▄●▂▁   #32 of 313     58%      8
coding       ▁▁▁▃▅▇█▇▅▄●▁   #22 of 469     31%      3
math         █····█·····●     #1 of 3*    1.9%      1
web search   █·······█··●     #1 of 3*    5.1%      2

agent tasks · #32 of 313 · 58% of its score

Trails [Grok 4.7](grok-4-7.md), [GLM-5.3](glm-5-3.md), [GPT-6 Luna](gpt-6-luna.md) here, though they rank lower overall.

test                                    result  vs the others in its run  difficulty            counts  why
Arena Agent (October 2026 aggrega…  0.0122 IPS  ▁▁▁▃▃▂●▁▂▂▁▁ beats 31/59  avg -0.379 IPS           30%  sharp
SWE-bench Pro (split unspecified)…       65.7%  ▂·▁▁▁▁▂▃█●▂▁ beats 51/58  very hard · avg 6.8%    6.7%  sharp · at its level · small run
Terminal-Bench 2.1 (maker)               85.4%  ▁·▁·▁▁▄▃▆█▅● beats 53/63  hard · avg 19%          4.9%  sharp · small run
APEX-Agents (pass@1) (maker)             37.1%  ▃▃··▃··█▃·▃● beats 8/8    very hard · avg 3.4%    4.4%  at its level · small run
SWE-bench Multilingual (maker)           82.9%  ▁··▁▃▁▅██●▂▁ beats 57/65  hard · avg 27%          3.9%  small run
MCP-Atlas (public set) (maker)           83.7%  ▂·▁▂▂·▁▃█▆▆● beats 39/39  hard · avg 32%          3.2%  small run
Humanity's Last Exam with tools (…       55.4%  ▁·▁▂▂▁▄▄▇●▄▅ beats 46/62  very hard · avg 14%     3.0%  at its level · small run
WideSearch (maker)                       83.9%  █·▂▆··▂···▂● beats 10/10  medium · avg 43%        2.1%  small run

coding · #22 of 469 · 31% of its score

Beats [GPT-5.6 Sol](gpt-5-6-sol.md), [GPT-5.5](gpt-5-5.md), [MiMo-V2.6-Pro](mimo-v2-6-pro.md) here, though they rank higher overall. Trails [Grok 4.7](grok-4-7.md), [Qwen3.8 Max (0902)](qwen3-8-max-0902.md), [Kimi K3](kimi-k3.md) here, though they rank lower overall.

test                                result  vs the others in its run    difficulty            counts  why
Arena WebDev (Code Arena)           1631.1  ▁▁▃▂▅█▃▆▅●▁▂ beats 109/129  avg 981                  20%  sharp
SWE-bench Pro (split unspecified)…   65.7%  ▂·▁▁▁▁▂▃█●▂▁ beats 51/58    very hard · avg 6.8%    6.7%  sharp · at its level · small run
SWE-bench Multilingual (maker)       82.9%  ▁··▁▃▁▅██●▂▁ beats 57/65    hard · avg 27%          3.9%  small run

math · #1 of 3 · 1.9% of its score

test                             result  vs the others in its run  difficulty           counts  why
MathArena Apex (pass@1) (maker)   74.2%  ▄····▄···▄█● beats 5/5    very hard · avg 14%    1.9%  at its level · small run

web search · #1 of 3 · 5.1% of its score

test                                result  vs the others in its run  difficulty           counts  why
Humanity's Last Exam with tools (…   55.4%  ▁·▁▂▂▁▄▄▇●▄▅ beats 46/62  very hard · avg 14%    3.0%  at its level · small run
WideSearch (maker)                   83.9%  █·▂▆··▂···▂● beats 10/10  medium · avg 43%       2.1%  small run

other tests (5)

test                                result  vs the others in its run    difficulty            counts  why
GPQA Diamond (maker)                 92.3%  ▁▁▁▁▂▄▅▅███● beats 297/322  hard · avg 40%          6.9%  sharp · maxed out · small run
Humanity's Last Exam (maker)         43.4%  ▁▁▁▄▆▆▇██●▁▃ beats 103/111  very hard · avg 5.1%    4.7%  at its level · small run
NL2Repo (maker)                      58.9%  ▂▂▂·▃▃▅█▃▂●▃ beats 19/22    very hard · avg 12%     3.4%  at its level · small run
SkillsBench 79-task text-only (Cl…   62.9%  █··········● beats 1/1      hard · avg 19%          2.4%  at its level · small run
SuperChem (maker)                    66.4%  █··········● beats 1/1      hard · avg 22%          2.0%  at its level · small run

## How the score was worked out Each result implies a score on its own, after adjusting for how hard that test is and who ran it. The model's score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its *counts* column above; here it is added up by who ran the tests. [Method](../method.md)

source                                    results  share of the score
LMArena leaderboard dataset                     2   50% ▇▇▇▇▇▇
Tencent Hy4 preview model card (submitt…       13   50% ▇▇▇▇▇▇
prior: pull toward the middle                        0% ▇
weighted average                                   = 76.7

_{Licences: LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Tencent Hy4 preview model card ([Apache-2.0](https://huggingface.co/tencent/Hy4-preview/blob/705d81ee51566a186d645b74c974d642ef2828fe/LICENSE))} ## How to run it ### Through an API api (Tencent Cloud TokenHub) · $1.25/M blended · $0.834 in / $2.501 out per 1M tokens · [cloud.tencent.com](https://cloud.tencent.com/document/product/1823/130051) _{[price source](https://www.tencentcloud.com/document/product/1300/78937): USD per 1M tokens, Tencent Cloud TokenHub (international site) model pricing, Hy4 preview, Singapore region tab: input $0.834, output $2.501; cache $0.042 excluded. The same rate appears under the other region tabs. The China site (cloud.tencent.com/document/product/1823/130055) lists CNY 6/18. Currency: USD. Tencent Cloud official pricing page (last updated 2026-10-03), observed 2026-10-07. · [context source](https://huggingface.co/tencent/Hy4-preview/blob/705d81ee51566a186d645b74c974d642ef2828fe/README.md): Context length stated on the official model card.} ### llama.cpp (recommended) 1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md) ```sh # macOS / Linux curl -LsSf https://llama.app/install.sh | sh # Windows irm https://llama.app/install.ps1 | iex ``` 2. Run Hy4 preview (770B-A49B) (IQ4_XS, needs ~417 GB, from [huggingface.co/6block](https://huggingface.co/6block/Hy4-preview-GGUF/tree/70f220329e9c6dbd9715f6e3a2b37b9f9291cf4e)) ```sh llama cli -hf 6block/Hy4-preview-GGUF:IQ4_XS ``` 3. Or start a local server and open the address it prints (http://localhost:9931) ```sh llama serve -hf 6block/Hy4-preview-GGUF:IQ4_XS ``` Other versions: IQ2_XS (~235.1 GB) `llama cli -hf 6block/Hy4-preview-GGUF:IQ2_XS` ### LM Studio 1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download) 2. Run Hy4 preview (770B-A49B) (IQ4_XS, needs ~417 GB, from [huggingface.co/6block](https://huggingface.co/6block/Hy4-preview-GGUF/tree/70f220329e9c6dbd9715f6e3a2b37b9f9291cf4e)) ```sh lms get https://huggingface.co/6block/Hy4-preview-GGUF@IQ4_XS -y ``` 3. Then run lms ls to see its name, and lms chat followed by that name Other versions: IQ2_XS (~235.1 GB) `lms get https://huggingface.co/6block/Hy4-preview-GGUF@IQ2_XS -y` ### Other ways to run it 1. Run it as published (fp8, needs ~813.77 GB, NVIDIA only, from [huggingface.co/tencent](https://huggingface.co/tencent/Hy4-preview/blob/705d81ee51566a186d645b74c974d642ef2828fe/README.md)) ```sh docker run --gpus all -p 8000:8000 --ipc=host -v ~/.cache/huggingface:/root/.cache/huggingface vllm/vllm-openai:hy4-preview tencent/Hy4-preview-FP8 --tensor-parallel-size 8 --speculative-config '{"num_speculative_tokens":3,"method":"mtp"}' --attention-backend FLASHMLA_SPARSE --tool-call-parser hy_v4 --reasoning-parser hy_v4 --enable-auto-tool-choice --port 8000 --served-model-name hy4-preview ``` Other versions: BF16 (~1559.98 GB, NVIDIA only) `vllm serve tencent/Hy4-preview` Model repository and documentation: [huggingface.co/tencent](https://huggingface.co/tencent/Hy4-preview) [interactive version](https://whatmodel.sh/model/hy4-preview) · [JSON](https://whatmodel.sh/api?format=json&all=1)