MiniMax-M2.1

+ contribute

MiniMax-M2.1 is an open-weights language model from MiniMax, released December 2025. Its published size is 228.7B parameters (10B active per token). Licence: other.

Catalogue identity notes

Comparison-only, proposed so StepFun's own reproductions in the Step 3.5 Flash card count (column 'MiniMax M2.1', '10B' activated / '230B' total). Weights: https://huggingface.co/MiniMaxAI/MiniMax-M2.1 (sha cd97f59135f37b2a6bf09356e485d5e4aeb7dc9c, created 2025-12-20, licence 'modified-mit', 228,689,764,864 parameters in safetensors metadata). Between catalogued minimax-m2 and minimax-m2-5.

65.7±1.2#135 of 719 LLMsfrom 12 benchmark results across 5 sources

$ how it compares

The LLMs ranked either side of MiniMax-M2.1, out of 828 with a score, then more from MiniMax.

#	model	score	where	RAM	released
152	DeepSeek V3.1 DeepSeek	66±1	api $0.42/M +local	~110GB	2025-08
153	Spark-X2.5 4B iFLYTEK (SparkLLM)	66±1	local	~3.9GB	2026-08
154	MiniMax-M2.1 MiniMax	66±1	local	~79GB	2025-12
155	Nemotron Cascade 2 30B-A3B NVIDIA	66±1	local	~20GB	2026-03
156	Laguna XS 2.1 (33B-A3B) Poolside	66±2	local	~21GB	2026-06
	more from MiniMax
84	MiniMax-M3 MiniMax	70±1	api $0.48/M +local	~137GB	2026-06
120	MiniMax-M2.7 MiniMax	68±1	api $0.53/M +local	~75GB	2026-03
150	MiniMax-M2.5 MiniMax	66±1	api $0.53/M +local	~75GB	2026-02

see all LLMs · find the ones that fit your machine

$ what it's good at

agent tasks▁▁▃▆█▆▅▅▅▄▃▁ #96 of 300 · 43% of its score · 4 tests

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

test	result	vs the others in its run	difficulty	counts	why
ContractBench success rate (33 observation-contract tool tasks)	60.6% ✎	▆··▆▄▂▄▂▆▆█▄ beats 16/23	very hard · avg 14%	▇▇▇▇▇ 13%	sharp · at its level
τ²-bench Telecom lab	87.0% ✎	▁·▅█▆▅▂▃▃▂▃▁ beats 65/89	medium · avg 35%	▇▇▇▇▇ 12%	sharp · small run
τ²-bench aggregate	86.6% ✎	▄······▄█·▄▄ beats 4/5	medium · avg 46%	▇▇▇▇ 10%	typical
BrowseComp-ZH	47.8% ✎	▁▁▁▁▁▃·▁▄▅▄█ beats 11/23	hard · avg 18%	▇▇▇▇ 9.4%	at its level

coding▁▁▁▃▆██▇▆▄▃▁ #96 of 458 * · 24% of its score · 1 test

Beats Grok 4.20 (0309 reasoning), Grok 4.3, Muse Glimmer 30B here, though they rank higher overall. Trails Llama 3.1 Nemotron Ultra 253B v1 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Arena WebDev (Code Arena)	1384.4 ✎	▁▂▄▃▅█▄▇▆▄▁▁ beats 38/119	avg 981	▇▇▇▇▇▇▇▇▇ 24%	sharp

chat & writing▁▁▁▃▅▄▅█▆▆▇▃ #125 of 393 * · 13% of its score · 1 test

Beats LongCat 2.0 (1.6T-A48B) here, though they rank higher overall. Trails LFM2.5 2.6B, Nemotron-Cascade-8B, DeepSeek V3 0324 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
LMArena (text)	1384.6 ✎	▁▁▁▂▃▃▃▆▆▇█▃ beats 224/379	avg 1268	▇▇▇▇▇ 13%	blunt

math▁▂▃▅▆▅▄█▄▁▁▁ #201 of 398 · 10% of its score · 3 tests

Trails Olmo 3 7B Instruct, Phi-4-mini-reasoning 3.8B, DeepSeek R1 Distill Qwen 7B here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
IMO-AnswerBench	60.4% ✎	▁▁·▁▁▃▄▅█▄▂▃ beats 5/42	medium · avg 35%	▇▇ 5.3%	at its level · small run
HMMT November 2025	74.3% ✎	▁▁····▁▅▃█▇▁ beats 3/23	medium · avg 65%	▇ 2.6%	noisy
HMMT February 2025	71.0% ✎	▁▁▁▃▁▅▅▃██▁▄ beats 9/31	medium · avg 58%	▇ 1.9%	noisy · small run

web search██···█····██ #5 of 5 * · 9.4% of its score · 1 test

test	result	vs the others in its run	difficulty	counts	why
BrowseComp-ZH	47.8% ✎	▁▁▁▁▁▃·▁▄▅▄█ beats 11/23	hard · avg 18%	▇▇▇▇ 9.4%	at its level

other tests3 not in a task · 10% of its score

test	result	vs the others in its run	difficulty	counts	why
ResearchRubrics	60.2% ✎	███··██·█··█ beats 4/6	hard · avg 31%	▇▇▇ 6.3%	at its level
GAIA (no-file subset)	64.3% ✎	▃▃····█▃···▃ beats 1/6	medium · avg 47%	▇ 3.3%	at its level
Vectara hallucination leaderboard (HHEM-2.3, factual consistency rate)	88.2% ✎	▁▂▁▂▅█▂▃▄▃▁▁ beats 27/98	very easy · avg 90%	▇ 0.0%	blunt

$ where to run it

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 MiniMax-M2.1 · needs ~138.3 GB · from huggingface.co/unsloth
chat in the terminal
llama cli -hf unsloth/MiniMax-M2.1-GGUF:Q4_K_M
Or start a local server and open the address it prints (http://localhost:9931)
llama serve -hf unsloth/MiniMax-M2.1-GGUF:Q4_K_M

LM Studio

version

Install LM Studio · Open the app once before using the lms command.
download it (macOS / Windows / Linux) ↗
Run MiniMax-M2.1 · needs ~138.3 GB · from huggingface.co/unsloth
open in LM Studio ↗
Or download it in the terminal
lms get https://huggingface.co/unsloth/MiniMax-M2.1-GGUF@Q4_K_M -y
Then run lms ls to see its name, and lms chat followed by that name

MLX (Mac)

version

Install MLX (Mac) · Apple silicon Macs only, text only.
with uv
uv tool install mlx-lm
or
pip install mlx-lm
Run MiniMax-M2.1 · needs ~128.7 GB, Mac only · from huggingface.co/mlx-community
chat in the terminal
mlx_lm.chat --model mlx-community/MiniMax-M2.1-4bit
Or start an OpenAI-compatible server at http://localhost:8080/v1
mlx_lm.server --model mlx-community/MiniMax-M2.1-4bit

Official weights and documentation: huggingface.co/MiniMaxAI ↗

$ how the score was worked out · stepfun-ai/Step-3.5-Flash mod… 38%, LMArena leaderboard dataset 37%, ContractBench: Can LLM Agents… 13% 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
stepfun-ai/Step-3.5-Flash model card	7	▇▇▇▇▇▇▇▇▇▇▇ 38%	Apache-2.0
LMArena leaderboard dataset	2	▇▇▇▇▇▇▇▇▇▇▇ 37%	CC BY 4.0
ContractBench: Can LLM Agents Preserve Observation Contracts? (arXiv:2605.17281v2)	1	▇▇▇▇ 13%	CC BY 4.0
MiniMax vendor-reported benchmark scores for its own models (Hugging Face model cards) · lab-reported	1	▇▇▇ 12%	Factual vendor-reported score (not licensed content)
Vectara hallucination leaderboard (vectara/hallucination-leaderboard repository)	1	▇ 0%	Apache-2.0
prior · pull toward the middle		▇ 1%
weighted average		= 65.7

$ 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