whatmodel.sh / Qwen3.8 Max (2.4T-A95B)

Qwen3.8 Max (2.4T-A95B)

+ contribute

Qwen3.8 Max (2.4T-A95B) is an open-weights language model from Alibaba, released August 2026. Its published size is 2400B parameters (95B active per token) and a 262k-token context window. Licence: qwen3.8-max-license.

78.2±0.6#18 of 719 LLMsfrom 23 benchmark results across 8 sources

$ how it compares

The LLMs ranked either side of Qwen3.8 Max (2.4T-A95B), out of 828 with a score, then more from Alibaba.

#	model	score	where	RAM	released
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
19	Claude Opus 4.8 Anthropic	78±0	api $10.00/M		2026-05
20	GPT-5.4 OpenAI	78±1	api $5.63/M		2026-03
	more from Alibaba
34	Qwen3.8 Max (0902) Alibaba	76±1	api $3.00/M		2026-09
37	Qwen3.8-Flash-Next Alibaba	75±1	local	~59GB	2026-08
38	Qwen3.8-Omni-Flash Alibaba	75±2	api $0.23/M		2026-09

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #18 of 300 · 13% of its score · 3 tests

Beats GPT-5.5, Hy4 preview (770B-A49B), GPT-5.6 Terra here, though they rank higher overall. Trails GLM-5.3, Claude Haiku 5.5, DeepSeek V4.1 Flash here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena Agent (October 2026 aggregate IPS)	0.0251 IPS ✎	▁▁▁▄▂▃▇█▂▂▁▃ beats 28/49	avg -0.312 IPS	▇▇▇ 7.1%	sharp
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)	44.4% ✎	█▄·█·▄·██▄·▄ beats 10/11	very hard · avg 4.7%	▇ 3.0%	at its level
τ³-bench banking_knowledge (v1.0.1 grading)	55.1% ✎	▃▃▂▃·▆▃▂█▅▅▃ beats 25/25	very hard · avg 4.1%	▇ 2.9%	at its level

coding▁▁▁▃▆██▇▆▄▃▁ #10 of 458 · 11% of its score · 3 tests

Beats GPT-5.6 Sol, Claude Fable 5, GPT-5.5 here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1671.2 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 111/119	avg 981	▇▇▇ 6.3%	sharp
GameLogicBench solve rate (72 Godot game-logic tasks, Claude Code scaffold)	44.4% ✎	█▄·█·▄·██▄·▄ beats 10/11	very hard · avg 4.7%	▇ 3.0%	at its level
Luau Leetcode fully solved (Torpedo Software)	77.4% ✎	▁▁▁▁·▁▂▃▅▃█▇ beats 27/49	hard · avg 18%	▇ 2.1%	typical

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #18 of 393 * · 3.6% of its score · 1 test

Beats GPT-6 Astra, Claude Sonnet 5.5, GPT-6 Sol here, though they rank higher overall. Trails Gemini 3.6 Flash, Claude Opus 4.6, Muse Spark 1.1 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1481.6 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 360/379	avg 1268	▇ 3.6%	blunt

math▁▂▃▅▆▅▄█▄▃▁▁ #19 of 398 · 54% of its score · 9 tests

Trails GPT-5.6 Luna, GPT-6 Luna, GPT-5.4 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
FrontierMath tier 4 v2	46.3% ✎	▁▂▁▂█▅▂▂▁▁▁▁ beats 45/63	very hard · avg 0.1%	▇▇▇▇▇▇▇ 17%	sharp · at its level
FrontierMath tiers 1–3 v2	74.7% ✎	▂··▁▁▃▄▄█▅▃▃ beats 64/82	very hard · avg 1.2%	▇▇▇▇▇ 13%	sharp
MathArena ArXivMath June 2026 (48 problems)	75.7% ✎	▂···▄·██▆▆▄▂ beats 13/19	very hard · avg 4.6%	▇▇▇ 8.1%	sharp · at its level
MathArena ArXivMath April 2026 (40 problems)	59.2% ✎	▁·····▅██▄·▁ beats 13/20	very hard · avg 7.2%	▇▇▇ 7.5%	sharp · at its level
MathArena BrokenArXiv April 2026 (61 problems)	45.5% ✎	▂▂▆█▃▃▂▂▂··▂ beats 13/18	very hard · avg 0.7%	▇ 3.6%	sharp · at its level · noisy
MathArena ArXivMath May 2026 (40 problems)	69.2% ✎	▁··▁▄█▄▁·▃·▁ beats 13/17	very hard · avg 10%	▇ 2.6%	at its level
MathArena ArXivMath August 2026, with tools (57 problems)	57.9% ✎	█▄·▄·······█ beats 3/5	very hard · avg 4.8%	▇ 1.8%	sharp · at its level
MathArena BrokenArXiv August 2026, with tools (56 problems)	69.6% ✎	▄▄·····▄·▄·█ beats 3/5	very hard · avg 5.5%	▇ 0.8%	at its level · noisy

1 more test

test	result	vs the others in its run	difficulty	counts	why
OTIS mock AIME 2024–2025	99.4% ✎	▃▃▄▃▃▃▇▇█▃▂▅ beats 183/193	very hard · avg 8.5%	▇ 0.4%	sharp · maxed out · noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #2 of 147 * · 5.5% of its score · 2 tests

Beats GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5.1 here, though they rank higher overall.

test	result	vs the others in its run	difficulty	counts	why
CharXiv (reasoning questions) lab	88.4% ✎	▄▂▄▆▄▄▆··▄█▃ beats 21/21	medium · avg 39%	▇ 3.1%	typical
Arena vision (style controlled)	1301.2 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 144/145	avg 1068	▇ 2.4%	blunt

other tests6 not in a task · 15% of its score

test	result	vs the others in its run	difficulty	counts	why
Mystery Game Puzzles (Epoch)	38.0% ✎	█··▂▃▇▄█▂▃▁▁ beats 61/74	very hard · avg 9.4%	▇▇ 6.1%	sharp · at its level · noisy
Chess Puzzles (Epoch)	29.0% ✎	█▁▁▁▁▂▁▄▃▂▁▁ beats 102/135	very hard · avg 5.9%	▇ 3.5%	sharp · at its level · noisy
GPQA Diamond	92.7% ✎	▁▁▁▁▂▄▆▄▅▇█▅ beats 190/208	hard · avg 40%	▇ 1.9%	sharp
SimpleQA Verified	45.8% ✎	▁▁▃▂▃▃▆▆█▃▂▅ beats 43/78	very hard · avg 8.0%	▇ 1.4%	at its level
Icelandic LLM Leaderboard average (Miðeind, Sept 2026 suite)	44.1% ✎	▁▁▃▂▂▄▁▃▄▄▄█ beats 16/38	very hard · avg 14%	▇ 1.2%	at its level
RobloxQA v2.0 multiple choice (Torpedo Software)	95.4% ✎	▁▁▃▁█▄▃▃▆▂▂▂ beats 29/49	easy · avg 78%	▇ 0.7%	maxed out

$ where to run it

Through an API

api · Deep Infra · $3.00/M · $2 in / $6 out per 1M tokens · 262k context · deepinfra.com ↗ · price sourceStandard base token rate from models.dev. Provider-specific tiers, long-context surcharges, caching and promotions may differ.

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 Qwen3.8 Max (2.4T-A95B) · needs ~1440.5 GB · from huggingface.co/unsloth
chat in the terminal
llama cli -hf unsloth/Qwen3.8-2.4T-A95B-GGUF:UD-IQ4_XS
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf unsloth/Qwen3.8-2.4T-A95B-GGUF:UD-IQ4_XS

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Qwen3.8 Max (2.4T-A95B) · needs ~1440.5 GB · from huggingface.co/unsloth
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/unsloth/Qwen3.8-2.4T-A95B-GGUF@UD-IQ4_XS -y
Then run lms ls to see its name, and lms chat followed by that name

Official weights and documentation: huggingface.co/Qwen ↗

$ how the score was worked out · Epoch AI — internal evaluatio… 43%, MathArena ArXivMath / BrokenA… 24%, LMArena leaderboard dataset 19% and 5 more

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
Epoch AI — internal evaluations	7	▇▇▇▇▇▇▇▇▇▇▇▇▇ 43%	CC BY 4.0
MathArena ArXivMath / BrokenArXiv / ArXivLean model-output datasets	6	▇▇▇▇▇▇▇ 24%	CC BY-SA 4.0
LMArena leaderboard dataset	4	▇▇▇▇▇▇ 19%	CC BY 4.0
Qwen (Alibaba) vendor-reported benchmark scores for its own models (qwen.ai launch posts, Qwen2.5-VL and Qwen3-VL technical reports, Qwen3.8-Flash-Next card) · lab-reported	1	▇ 3%	Factual vendor-reported score (not licensed content)
GameLogicBench: Evaluating Coding Agents on Runtime Game Logic with Tick-Level State Assertions (arXiv:2609.21562v2)	1	▇ 3%	CC BY 4.0
τ²/τ³-bench leaderboard: Sierra's own evaluation runs	1	▇ 3%	MIT
Roblox LLM Leaderboard (Torpedo Software): RobloxQA v2.0 and Luau Leetcode	2	▇ 3%	MIT
Miðeind Icelandic LLM Leaderboard	1	▇ 1%	CC BY 4.0
prior · pull toward the middle		▇ 0%
weighted average		= 78.2