whatmodel.sh / Mistral Medium 3.5 (128B)

Mistral Medium 3.5 (128B)

+ contribute

Mistral Medium 3.5 (128B) is an open-weights language model from Mistral, released April 2026. Its published size is 128B parameters and a 262k-token context window. Licence: modified-mit.

64.5±0.9#152 of 719 LLMsfrom 22 benchmark results across 5 sources

$ how it compares

The LLMs ranked either side of Mistral Medium 3.5 (128B), out of 828 with a score, then more from Mistral.

#	model	score	where	RAM	released
172	Laguna XS.2 Poolside	65±2	local	~21GB	2026-04
173	Qwen3 235B-A22B Instruct 2507 Alibaba	65±1	api $0.20/M +local	~76GB	2025-07
174	Mistral Medium 3.5 (128B) Mistral	65±1	api $3.00/M +local	~42GB	2026-04
175	Command A+ (05-2026) Cohere	64±1	api price unknown +local	~71GB	2026-05
176	Claude Haiku 4.5 Anthropic	64±1	api $2.00/M		2025-10
	more from Mistral
79	Mistral Large 4 (public preview) Mistral	71±2	api $2.06/M		2026-10
180	Mistral Medium 3.1 (2508) Mistral	64±2	comparison only		2025-08
200	Magistral Medium 1.2 (2509) Mistral	63±2	comparison only		2025-09

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #103 of 300 · 28% of its score · 8 tests

Beats Grok 4.20 (0309 reasoning), Grok 4.3, o3-pro here, though they rank higher overall. Trails SmolLM3 3B, Nova 2 Pro (Preview) here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena Agent (October 2026 aggregate IPS)	-0.1238 IPS ✎	▁▃▁▄▂▃▇█▂▂▁▃ beats 1/49	avg -0.312 IPS	▇▇▇▇▇▇ 14%	sharp
Terminal-Bench Hard	33.3% ✎	▄········██▄ beats 3/5	very hard · avg 4.5%	▇ 3.2%	sharp
APEX-Agents (pass@1)	6.1% ✎	▄▄····▄···█▄ beats 2/5	very hard · avg 1.8%	▇ 3.0%	too hard for it
MCP-Atlas (public set)	30.7% ✎	▄▄▄·······█▄ beats 1/5	hard · avg 19%	▇ 2.6%	at its level
EEGAgentBench overall (tool-based EEG analysis agents, 6 tasks)	41.9% ✎	▂·▁▁··▇▇█▂·▁ beats 4/25	hard · avg 29%	▇ 2.0%	blunt · at its level
SWE-bench Verified	69.6% ✎	▄··········█ beats 2/5	hard · avg 16%	▇ 1.9%	sharp · at its level · noisy
τ³-bench Banking	5.8% ✎	█·▄▄······▄▄ beats 0/5	very hard · avg 5.1%	▇ 0.8%	blunt
τ³-bench telecom (tau2-bench ≥ v1.0.1) lab	91.4% ✎	▄·▄█▂··▂▂·▂▄ beats 6/13	medium · avg 62%	—	not counted yet: no other model in the same run to compare it with

coding▁▁▁▃▆██▇▆▄▃▁ #148 of 458 · 18% of its score · 3 tests

Beats Gemini 2.5 Pro, Gemini 3.1 Flash-Lite, Qwen3-Coder-Next (80B-A3B) here, though they rank higher overall. Trails Ministral 3 8B Reasoning 2512, Phi-4-reasoning 14B, DeepSeek V3 0324 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1262.9 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 16/119	avg 981	▇▇▇▇▇ 12%	sharp
LiveCodeBench v6 (window unspecified)	84.9% ✎	▄······▄▄▄·█ beats 1/5	medium · avg 50%	▇ 3.2%	typical
SWE-bench Verified	69.6% ✎	▄··········█ beats 2/5	hard · avg 16%	▇ 1.9%	sharp · at its level · noisy

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #80 of 393 * · 7.0% of its score · 1 test

Beats LongCat 2.0 (1.6T-A48B), Inkling Small (266B-A12B), Muse Glimmer 30B here, though they rank higher overall. Trails Gemma 4 12B, Gemma 4 26B-A4B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1426.9 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 282/379	avg 1268	▇▇▇ 7.0%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #175 of 398 * · 9.5% of its score · 2 tests

Beats MiniMax-M2.1 here, though they rank higher overall. Trails OpenThinker3 7B, Olmo 3 7B Think, OpenThinker2 32B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
AIME 2026	89.0% ✎	▄▄··▄····█·▄ beats 1/5	medium · avg 64%	▇▇▇ 6.5%	sharp
HMMT February 2026	62.9% ✎	█▄▄········█ beats 1/5	hard · avg 31%	▇ 2.9%	sharp · at its level · noisy

vision & documents▁▁▃▂▁▃▄▂▆▇█▄ #75 of 147 * · 4.8% of its score · 1 test

Beats DeepSeek V4 Flash Vision (Exp), GPT-5.4 nano, GPT-5 mini here, though they rank higher overall. Trails GPT-4.1, Gemma 4 12B, Qwen3.5 9B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena vision (style controlled)	1199.5 ✎	▁▁▂▂▂▂▄▅▅▆██ beats 70/145	avg 1068	▇▇ 4.8%	blunt

other tests9 not in a task · 35% of its score

test	result	vs the others in its run	difficulty	counts	why
GPQA Diamond	77.5% ✎	█···██·█··██ beats 2/5	hard · avg 42%	▇▇▇▇▇▇ 15%	sharp · at its level
MMLU-Pro	81.2% ✎	▄·▄▄····▄··█ beats 2/5	medium · avg 55%	▇▇ 4.9%	typical
Humanity's Last Exam	12.8% ✎	█▄······▄·▄▄ beats 2/5	very hard · avg 4.7%	▇▇ 4.8%	typical
IFBench (aggregation unspecified)	69.0% ✎	▄···▄▄··▄··█ beats 2/5	medium · avg 43%	▇ 3.2%	at its level
MultiChallenge	49.8% ✎	█··▄·▄·····█ beats 3/5	hard · avg 25%	▇ 2.6%	at its level
Icelandic LLM Leaderboard average (Miðeind, Sept 2026 suite)	14.9% ✎	▁▁▃▂▂▄▁▃▄▄▄█ beats 3/38	very hard · avg 14%	▇ 2.2%	at its level
WMT25 Terminology Translation (term success rate)	90.3% ✎	▂·▂··▂·▄▂·█▆ beats 9/12	medium · avg 64%	▇ 1.3%	noisy
WMT24++	76.7% ✎	█·▃·▅▅▃···▃▃ beats 1/10	easy · avg 71%	▇ 0.6%	blunt

1 more test

test	result	vs the others in its run	difficulty	counts	why
WMT26 General MT (GEMBA-ESA, LLM judge)	75.7% ✎	▄·▄▄▄·▄██▄█▄ beats 4/12	easy · avg 65%	▇ 0.5%	blunt · at its level

$ where to run it

Through an API

api · Mistral AI · $3.00/M · $1.5 in / $7.5 out per 1M tokens · 262k context · docs.mistral.ai ↗ · price sourceUSD per 1M text tokens, standard uncached input and output. Product: La Plateforme chat completions. API model: mistral-medium-3-5. Settings: service tier=standard, batch=no, cache=uncached input. Additional/excluded charges: Tools, cache storage, priority service and other modalities excluded.. Currency: USD.

llama.cpp · recommended

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 Mistral Medium 3.5 (128B) · needs ~38.9 GB · from huggingface.co/unsloth
chat in the terminal
llama cli -hf unsloth/Mistral-Medium-3.5-128B-GGUF:UD-Q2_K_XL
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf unsloth/Mistral-Medium-3.5-128B-GGUF:UD-Q2_K_XL

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run Mistral Medium 3.5 (128B) · needs ~75.4 GB, Mac only · from huggingface.co/froggeric
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/froggeric/Mistral-Medium-3.5-128B-MLX-4bit -y
Then run lms ls to see its name, and lms chat followed by that name

Ollama

Install Ollama
download it (macOS / Windows) ↗
Linux
curl -fsSL https://ollama.com/install.sh | sh
Run Mistral Medium 3.5 (128B) · needs ~77.3 GB · from ollama.com
chat in the terminal
ollama run mistral-medium-3.5:128b

MLX (Mac)

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run Mistral Medium 3.5 (128B) · needs ~75.4 GB, Mac only · from huggingface.co/froggeric
chat in the terminal
mlx_lm.chat --model froggeric/Mistral-Medium-3.5-128B-MLX-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model froggeric/Mistral-Medium-3.5-128B-MLX-4bit

Weights and documentation: huggingface.co/mistralai ↗

$ how the score was worked out · Solar Open 2 Technical Report… 55%, LMArena leaderboard dataset 38%, Cohere — North Small Translat… 2% and 2 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
Solar Open 2 Technical Report (Upstage, arXiv:2607.20062v2)	13	▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇▇ 55%	CC-BY-4.0
LMArena leaderboard dataset	4	▇▇▇▇▇▇▇▇▇▇▇ 38%	CC BY 4.0
Cohere — North Small Translate (arXiv:2609.13916v1)	3	▇ 2%	CC BY 4.0
Miðeind Icelandic LLM Leaderboard	1	▇ 2%	CC BY 4.0
EEGAgentBench: Benchmarking LLM Agents on Short- and Long-Horizon EEG Analysis (arXiv:2609.31632v1)	1	▇ 2%	CC BY 4.0
prior · pull toward the middle		▇ 0%
weighted average		= 64.5

not counted (1)

Mistral AI vendor-reported benchmark scores for its own models (Hugging Face model cards)

τ³-bench telecom (tau2-bench ≥ v1.0.1) [correct]

91.4%

fewer than three models have this test yet, or this source ran it for this model alone

$ 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