whatmodel.sh / Hy4 preview (770B-A49B)

Hy4 preview (770B-A49B)

+ contribute

Hy4 preview (770B-A49B) is an open-weights language model from Tencent, released August 2026. Its published size is 770B parameters (49B active per token) and a 1M-token context window. Licence: apache-2.0.

Catalogue identity notes

Arena entry 'hy4-preview' (webdev). Maker repo tencent/Hy4-preview @705d81e: licence tag apache-2.0. Parameters: card: '770B total parameters, of which 49B are activated per token' (backbone, excluding the 10B MTP layer). Released 2026-08: Tencent TokenHub changelog: 2026-08-28 'added Hy4 preview'. Arena name equals the API id hy4-preview; the card's own serving examples use served-model-name 'hy4-preview'.

79.1±1.1#16 of 719 LLMsfrom 15 benchmark results across 2 sources

$ how it compares

The LLMs ranked either side of Hy4 preview (770B-A49B), out of 828 with a score, then more from Tencent.

#	model	score	where	RAM	released
14	GPT-5.5 OpenAI	80±1	api $11.25/M		2026-04
15	MiMo-V2.6-Pro Xiaomi	79±1	api $0.54/M +local	~443GB	2026-09
16	Hy4 preview (770B-A49B) Tencent	79±1	api $1.25/M +local	~248GB	2026-08
17	GPT-5.6 Terra OpenAI	79±1	api $4.50/M		2026-07
18	Qwen3.8 Max (2.4T-A95B) Alibaba	78±1	api $3.00/M +local	~380GB	2026-08
	more from Tencent
55	Hunyuan Hy3 (295B-A21B) Tencent	74±1	api $0.23/M +local	~96GB	2026-07
123	Hy3 preview (295B-A21B) Tencent	68±1	local	~629GB	2026-04

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #21 of 300 · 55% of its score · 8 tests

Beats GPT-5.5 here, though they rank higher overall. Trails Grok 4.7, GLM-5.3, Claude Haiku 5.5 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena Agent (October 2026 aggregate IPS)	0.0395 IPS ✎	▁▁▁▄▂▃▇█▂▂▁▃ beats 32/49	avg -0.312 IPS	▇▇▇▇▇▇▇▇▇ 23%	sharp
SWE-bench Pro (split unspecified) lab	65.7% ✎	▃·▁▁▁▂▂▃█▇▃▁ beats 48/55	very hard · avg 6.8%	▇▇▇ 7.0%	sharp · at its level · small run
Terminal-Bench 2.1 lab	85.4% ✎	▁·▁·▁▁▄▃▄█▃▇ beats 47/56	hard · avg 17%	▇▇ 5.7%	small run
APEX-Agents (pass@1) lab	37.1% ✎	▃▃··▃··█▃·▃▃ beats 8/8	very hard · avg 2.8%	▇▇ 5.4%	at its level · small run
SWE-bench Multilingual lab	82.9% ✎	▁··▁▃▂▅▇█▃▃▁ beats 52/60	hard · avg 27%	▇▇ 4.2%	small run
MCP-Atlas (public set) lab	83.7% ✎	▂·▁▂▁·▁▃█▆▆▃ beats 37/37	hard · avg 29%	▇ 3.7%	small run
Humanity's Last Exam with tools lab	55.4% ✎	▂▂▂▁▄▂▃▆█▇▂▆ beats 40/54	very hard · avg 15%	▇ 3.1%	at its level · small run
WideSearch lab	83.9% ✎	█·▂▆··▂···▂▃ beats 10/10	medium · avg 40%	▇ 2.5%	small run

coding▁▁▁▃▆██▇▆▄▃▁ #18 of 458 · 32% of its score · 3 tests

Beats GPT-5.6 Sol, GPT-5.5, MiMo-V2.6-Pro here, though they rank higher overall. Trails Grok 4.7, Qwen3.8-Flash-Next, Qwen3.8 Max (0902) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1633.2 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 104/119	avg 981	▇▇▇▇▇▇▇▇ 21%	sharp
SWE-bench Pro (split unspecified) lab	65.7% ✎	▃·▁▁▁▂▂▃█▇▃▁ beats 48/55	very hard · avg 6.8%	▇▇▇ 7.0%	sharp · at its level · small run
SWE-bench Multilingual lab	82.9% ✎	▁··▁▃▂▅▇█▃▃▁ beats 52/60	hard · avg 27%	▇▇ 4.2%	small run

math█····█·····█ #1 of 3 * · 2.5% of its score · 1 test

test	result	vs the others in its run	difficulty	counts	why
MathArena Apex (pass@1) lab	74.2% ✎	▄····▄···▄█▄ beats 5/5	very hard · avg 12%	▇ 2.5%	at its level · small run

web search█·······█··█ #1 of 3 * · 5.6% of its score · 2 tests

test	result	vs the others in its run	difficulty	counts	why
Humanity's Last Exam with tools lab	55.4% ✎	▂▂▂▁▄▂▃▆█▇▂▆ beats 40/54	very hard · avg 15%	▇ 3.1%	at its level · small run
WideSearch lab	83.9% ✎	█·▂▆··▂···▂▃ beats 10/10	medium · avg 40%	▇ 2.5%	small run

other tests5 not in a task · 22% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond lab	92.3% ✎	▁▁▁▁▂▄▅▅███▄ beats 287/310	hard · avg 40%	▇▇▇ 6.8%	sharp · maxed out · small run
Humanity's Last Exam lab	43.4% ✎	▁▁▁▃▆▆▇██▆▁▃ beats 96/104	very hard · avg 4.9%	▇▇ 5.3%	at its level · small run
NL2Repo lab	58.9% ✎	▂▂·▂▂▄█▆▆▂▄▆ beats 19/21	very hard · avg 10%	▇▇ 4.3%	at its level · small run
SkillsBench 79-task text-only (Claude Code, avg of 3) lab	62.9% ✎	█··········█ beats 1/1	hard · avg 16%	▇ 2.8%	at its level · small run
SuperChem lab	66.4% ✎	█··········█ beats 1/1	hard · avg 20%	▇ 2.4%	at its level · small run

$ where to run it

Through an API

api · Tencent Cloud TokenHub · $1.25/M · $0.834 in / $2.501 out per 1M tokens · 1M context · cloud.tencent.com ↗ · price sourceUSD 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.

llama.cpp · recommended

version

Install llama.cpp · Homebrew and winget installs name the commands llama-cli and llama-server instead.
macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
Windows
irm https://llama.app/install.ps1 | iex
other ways to install ↗
Run Hy4 preview (770B-A49B) · needs ~417 GB · from huggingface.co/6block
chat in the terminal
llama cli -hf 6block/Hy4-preview-GGUF:IQ4_XS
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf 6block/Hy4-preview-GGUF:IQ4_XS

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Hy4 preview (770B-A49B) · needs ~417 GB · from huggingface.co/6block
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/6block/Hy4-preview-GGUF@IQ4_XS -y
Then run lms ls to see its name, and lms chat followed by that name

Other ways to run it

version

Run it as published · needs ~813.77 GB, NVIDIA only · from huggingface.co/tencent (official)
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

Official weights and documentation: huggingface.co/tencent ↗

$ how the score was worked out · Tencent Hy4 preview model card 56%, LMArena leaderboard dataset 44%

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The 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's added up by who ran the tests. How it works

source	results	share of the score	licence
Tencent Hy4 preview model card · lab-reported	13	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 56%	Apache-2.0
LMArena leaderboard dataset	2	▇▇▇▇▇▇▇▇▇▇▇▇▇ 44%	CC BY 4.0
prior · pull toward the middle		▇ 1%
weighted average		= 79.1

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json