MiniMax-M2.1

MiniMaxAI MiniMax (Modified MIT) (modified-mit) out 10 months ago

229B parameters (8 of 256 experts per token), text in, text out, 192K tokens of context. 2 providers serve it from $0.30 per million input tokens.

Verified 2026-10-25 Rank #154 in LLMs · score 92.3 method v1 Sources 4 Downloads 30d 12.1k · #496 of 619 tracked

Can I run it?137GB at INT4Estimatefits 2×80 GBmethodCommunity: MiniMax M2.1: Built for Real-World Complex Tasks, Multi-Language Programming →

What does it cost?$0.30/1M incheapest of 1Minimaxfp82026-10-24

Can I use it?MiniMax (Modified MIT) (modified-mit)MIT with added terms; read the repo LICENSE before commercial hostinglicence file2026-10-24

Do people use it?12.1kdownloads, 30 days#496 of 619 we track · below the median+5 likes, downloads +14% this week1,374 likesHugging Face2026-10-24

Worth reading

Picked by rule from 12 tracked pieces: the lab's own post, the most shared, a field report, the latest.

MiniMax M2.1 used batch-based execution where tasks are grouped by dependency level.

From Open-Weight Models Are Getting Serious: GLM 4.7 vs. MiniMax M2.1 · blog.kilo.ai · 2026-01-09

The lab's own post

MiniMax M2.1: Built for Real-World Complex Tasks, Multi-Language Programming

MiniMax M2.1 is an open model designed for complex real-world tasks, featuring systematic Interleaved Thinking and improved multi-language support.

minimaxi.com 2025-12-23 228 HN pts · 81 comments 380 words sells inference

Most discussed · 16 HN pts

Open-Weight Models Are Getting Serious: GLM 4.7 vs. MiniMax M2.1

GLM 4.7 and MiniMax M2.1 successfully built a CLI task runner with 20 features in a multi-phase test.

MiniMax M2.1 used batch-based execution where tasks are grouped by dependency level.
GLM 4.7 implemented argument parsing manually to avoid external dependencies.

blog.kilo.ai 2026-01-09 16 HN pts 1.5k words

Latest · 9 months ago

MiniMax M2.1: Post-Training Experience and Insights for Agent Models - MiniMax News | MiniMax

MiniMax M2.1 is an open-source agent model with a Mixture-of-Experts architecture featuring 230B total parameters and ~10B activated parameters.

Environment construction requires expert knowledge to optimize execution flow when automated builds fail for specific languages or library versions.
The Forge framework supports integrating black-box Agents as binary executables by redirecting their base URL to an internal inference engine service.

minimax.io 2026-01-22 2.5k words sells inference

12 best of 12 tracked

The 4 picks above, plus these 8. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleTraining MiniMax-M2.1minimaxi.com · 2026-01-05 · 3 HN pts · official
ArticleMiniMax M2.1 Bets That ‘Most Usable’ Beats ‘Most Massive’hackernoon.com · undated
ArticleMiniMax M2.1: Scale and efficiency for multi-language developmenttessl.io · 2026-01-05
ArticleMiniMax M2.1: Architecture, Benchmarks, and Practical Deployment | AiCybr Blogaicybr.com · 2026-01-04
ArticleMiniMax M2.1: Open-Source MoE Model Sets Coding Benchmark - AI CERTs Newsaicerts.ai · 2026-01-02
ArticleMiniMax M2.1 Review: Fast, Cheap Coding AI That Actually Shipsquickleap.hashnode.dev · 2025-12-29
ArticleAn Analytical Review of MiniMax M2.1 | by Barnacle Goose | Mediummedium.com · 2025-12-23
ArticleMiniMax M2.1 Guide: Digital Employee for AI Codingdigitalapplied.com · 2025-12-23

Show all 8

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

#	Cheapest 2	In /1M	Out /1M	Context	Quant	Seen
1	Minimax	$0.30	$1.20	200K	fp8	Oct 24
2	Featherless	—	—	—	—	Oct 25

All 2 providers →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Served by 2 API providers; 37 quantized builds for local use.

As a service2Featherless, Minimax
On own hardware37quantized builds
Built on1312 fine-tunes · 1 adapters · 0 merges
In apps100+Hugging Face Spaces

Top Spaces running it: enzostvs/deepsite ♥ 16,604 · akhaliq/anycoder ♥ 3,321 · hadadxyz/ai ♥ 819 · pliny-the-prompter/obliteratus ♥ 621

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits two 80 GB GPUs at INT4; too big for one

About 137 GB at INT4, 274 GB at FP8, 549 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

9 dated articles and 30 repo commits so far

Teal is the weights (2025-12-20), amber triangles are 9 dated articles (each links to it), grey ticks are repo commits.

What people built on it

50 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

50 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF unsloth/MiniMax-M2.1-GGUF for llama.cpp, Ollama, LM Studio 252.8k downloads MLX mlx-community/MiniMax-M2.1-3bit for Apple silicon 87.4k downloads AWQ cyankiwi/MiniMax-M2.1-AWQ-4bit for vLLM, SGLang 100.5k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

3 rows

Benchmarks 3 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (3)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported SWE-bench · Write code with agents	49.4	—	model author	2026-10-07
author-reported SWE-bench Multilingual · Write code with agents	72.5	—	model author	2026-10-07
author-reported SWE-bench Verified · Write code with agents	74	—	model author	2026-10-07

Use-case evidence

5 use cases

What it is good at 5 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

22 fields

Listed in LLMs · MoE

Facts

MiniMax-M2.1 has 229B parameters. (safetensors metadata, observed 2026-10-24)
MiniMax-M2.1 is a mixture-of-experts model with 256 experts, routing 8 per token. (config.json, observed 2026-10-24)
Context length is 196,608 tokens (192K). (config.json max_position_embeddings, observed 2026-10-24)
License: MiniMax (Modified MIT) (modified-mit) (friction: review). MIT with added terms; read the repo LICENSE before commercial hosting. (license file, observed 2026-10-24)
Created 2025-12-20, last modified 2026-02-13. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	MiniMaxM2ForCausalLM
Model type	minimax_m2
Params by dtype	BF16: 1,230,021,632 · F32: 48,774,656 · F8_E4M3: 227,410,968,576
Size band	XXL
MoE	yes
Experts	256
Experts per token	8
Context length	196,608 (config.json max_position_embeddings)
Papers	arXiv:2509.06501
License	MiniMax (Modified MIT) (modified-mit) (other)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · MiniMaxM2ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · MiniMaxM2ForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for MiniMax-M2.1: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	549 GB · Multi-GPU node (<640 GB)
FP8 (1 byte/param)	274 GB · Multi-GPU node (<640 GB)
INT4 (0.5 bytes/param)	137 GB · Multi-GPU node (<640 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

40 variants

Lineage

Original release (shipped quantized by the lab) Family: MiniMaxAI/MiniMax-M2.1 50 derivatives on the Hub

quantized: 37finetune: 12adapter: 1

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
unsloth/MiniMax-M2.1-GGUF	quantized	GGUF	—	—	other	unsloth	195	252.8k
cerebras/MiniMax-M2.1-REAP-139B-A10B	quantized	—	139B	XXL	other	cerebras	32	193.3k
mradermacher/MiniMax-M2.1-i1-GGUF	quantized	GGUF	—	—	other	mradermacher	4	155k
cyankiwi/MiniMax-M2.1-AWQ-4bit	quantized	AWQ	36.8B	XL	other	cyankiwi	18	100.5k
mlx-community/MiniMax-M2.1-3bit	quantized	—	229B	XXL	other	mlx-community	5	87.4k
mlx-community/MiniMax-M2.1-8bit	quantized	-8bit	229B	XXL	other	mlx-community	2	83.9k
lukealonso/MiniMax-M2.1-NVFP4	quantized	NVFP4	115B	XXL	—	lukealonso	24	81.9k
Ex0bit/MiniMax-M2.1-PRISM	quantized	GGUF	—	—	mit	Ex0bit	109	49.3k
mlx-community/MiniMax-M2.1-4bit	quantized	-4bit	229B	XXL	other	mlx-community	5	39.4k
QuantTrio/MiniMax-M2.1-AWQ	quantized	AWQ	229B	XXL	other	QuantTrio	10	38.6k
amd/MiniMax-M2.1-MXFP4	quantized	MXFP4	116B	XXL	other	amd	1	31k
bartowski/MiniMaxAI_MiniMax-M2.1-GGUF	quantized	GGUF	—	—	other	bartowski	11	25.7k
mratsim/MiniMax-M2.1-FP8-INT4-AWQ	quantized	-FP8	39.1B	XL	other	mratsim	39	8k
AaryanK/MiniMax-M2.1-GGUF	quantized	GGUF	—	—	other	AaryanK	16	7.5k
cerebras/MiniMax-M2.1-REAP-172B-A10B	quantized	—	173B	XXL	other	cerebras	10	7.3k
mradermacher/MiniMax-M2.1-GGUF	quantized	GGUF	—	—	other	mradermacher	2	5.7k
GadflyII/MiniMax-M2.1-NVFP4	quantized	NVFP4	129B	XXL	mit	GadflyII	7	5.4k
mratsim/MiniMax-M2.1-BF16-INT4-AWQ	quantized	-BF16	39.1B	XL	other	mratsim	7	5.3k
plezan/MiniMax-M2.1-REAP-50-W4A16	quantized	-W4A16	16.7B	L	apache-2.0	plezan	5	4.8k
DevQuasar/MiniMaxAI.MiniMax-M2.1-GGUF	quantized	GGUF	—	—	—	DevQuasar	0	4.1k
John1604/MiniMax-M2.1-gguf	quantized	gguf	—	—	other	John1604	1	3.9k
mlx-community/MiniMax-M2.1-6bit	quantized	—	229B	XXL	other	mlx-community	2	3.3k
mlx-community/MiniMax-M2.1-5bit	quantized	—	229B	XXL	other	mlx-community	1	2.6k
unsloth/MiniMax-M2.1	quantized	—	229B	XXL	other	unsloth	13	1.2k
0xSero/MiniMax-M2.1-139B	quantized	—	139B	XXL	apache-2.0	0xSero	10	856
moxin-org/MiniMax-M2.1-GGUF	quantized	GGUF	—	—	mit	moxin-org	7	860
0xSero/MiniMax-M2.1-162B	quantized	—	162B	XXL	apache-2.0	0xSero	4	759
mlx-community/MiniMax-M2.1-mxfp4	quantized	mxfp4	229B	XXL	other	mlx-community	0	837
rushyrush/MiniMax-M2.1-REAP-172B-A10B-MXFP4_MOE-GGUF	quantized	GGUF	—	—	other	rushyrush	4	560
Felladrin/gguf-Q2_K_S-Mixed-AutoRound-MiniMax-M2.1	quantized	gguf	—	—	other	Felladrin	5	411
bullerwins/MiniMax-M2.1-GGUF	quantized	GGUF	—	—	other	bullerwins	2	455
anikifoss/MiniMax-M2.1-HQ4_K-GGUF	quantized	GGUF	—	—	mit	anikifoss	2	454
a2s-ai/MiniMax-M2.1-AWQ	quantized	AWQ	229B	XXL	other	a2s-ai	0	469
Firworks/MiniMax-M2.1-nvfp4	quantized	nvfp4	129B	XXL	—	Firworks	3	249
Doctor-Shotgun/MiniMax-M2.1-GGUF	quantized	GGUF	—	—	—	Doctor-Shotgun	2	189
TacGibs/MiniMaxM2.1-EXL3-3.04bpw	quantized	EXL3	44.4B	XL	—	TacGibs	2	104
INC4AI/MiniMax-M2.1-int4-AutoRound	quantized	int4	1.23B	S	—	INC4AI	1	135
KArina-2202/teste	finetune	—	67.0M	S	mit	KArina-2202	1	56
kunarmy99/minimaxAI-openRAIL	finetune	—	—	—	openrail	kunarmy99	2	0
chihqqqwe/aichi	finetune	—	—	—	apache-2.0	chihqqqwe	1	0

Repository changelog

30 commits

Repository changelog 30 commits

Date	Commit	SHA
2026-02-13	Update README.md (#28)	cd97f59
2026-01-28	Clarify supported context length in docs (#21)	69af314
2026-01-23	add mlx (#8)	927ea2b
2025-12-27	update bench.png (#7)	17f852d
2025-12-26	update	2713e50
2025-12-26	update	1ccbe32
2025-12-23	Update README.md	8d70a8a
2025-12-23	Update README.md	2bdeb8c
2025-12-23	Update chat_template.jinja	c6b717a
2025-12-23	Upload bench.png	20f0649
2025-12-23	update: README	1aeff0e
2025-12-22	update: use nightly sglang	0469488
2025-12-22	update: add toolcall docs	5b50d0e
2025-12-22	udpate: guide use nightly vllm	a423be8
2025-12-22	Create generation_config.json	2e162cb
2025-12-22	update: update model_identity in chat_template.jinja	7aef555
2025-12-21	update: support transformers	e551b49
2025-12-21	update: add config.json	500a52e
2025-12-21	update: vllm & sglang deploy guide	4d0619f
2025-12-21	Add files using upload-large-folder tool	3cc80b6
2025-12-21	Add files using upload-large-folder tool	dbc7874
2025-12-21	Add files using upload-large-folder tool	551920a
2025-12-21	Add files using upload-large-folder tool	6b3b637
2025-12-21	Add files using upload-large-folder tool	c0be2d6
2025-12-21	Add files using upload-large-folder tool	7c3ece6
2025-12-21	Add files using upload-large-folder tool	8fdad3a
2025-12-21	Add files using upload-large-folder tool	13a477f
2025-12-21	Add files using upload-large-folder tool	b91c862
2025-12-21	Add files using upload-large-folder tool	3f481ea
2025-12-21	Add files using upload-large-folder tool	de64fb3

All providers

2 providers · 2 endpoints

Where to run 2 providers · 2 endpoints

Hosted endpoints observed for MiniMax-M2.1, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.30/M.

Provider	Model ref	Context	Quant	$/M in	$/M out	tok/s	TTFT	Tools	Status	Observed
Minimax openrouter	minimax/minimax-m2.1	200K	fp8	$0.30	$1.20	—	—	yes	live	2026-10-24
Featherless hf	MiniMaxAI/MiniMax-M2.1	—	—	—	—	—	—	—	live	2026-10-25

Articles and coverage

12 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official3

MiniMax M2.1: Built for Real-World Complex Tasks, Multi-Language Programming

minimaxi.com2025-12-23officialHN front pageHN 228 pts · 81 comments380 words

MiniMax M2.1 is an open model designed for complex real-world tasks, featuring systematic Interleaved Thinking and improved multi-language support. It achieves an average score of 88.6 on the VIBE benchmark, demonstrating near-Claude Opus 4.5 full-stack capabilities. The release includes a faster M2.1-lightning API variant and automatic caching to reduce token consumption.written, reviewed, cited

MiniMax M2.1: Post-Training Experience and Insights for Agent Models - MiniMax News | MiniMax

minimax.io2026-01-22officialauthority2,460 words

MiniMax M2.1 is an open-source agent model with a Mixture-of-Experts architecture featuring 230B total parameters and ~10B activated parameters. It uses SWE Scaling to process 10,000+ runnable PRs creating 140,000+ variable tasks and incorporates AppDev with expert-in-the-loop specialists for full-stack development. The model also employs WebExplorer for long-horizon search, evolving queries to increase average reasoning turns from 7.9 to 9.9.written, reviewed, cited

Environment construction requires expert knowledge to optimize execution flow when automated builds fail for specific languages or library versions.
The Forge framework supports integrating black-box Agents as binary executables by redirecting their base URL to an internal inference engine service.
CISPO applies clipping to importance sampling weights to control gradient magnitude and prevent tokens from permanently losing gradients during training.

Training MiniMax-M2.1

minimaxi.com2026-01-05officialHN 3 pts · 0 comments

Analysis7

Open-Weight Models Are Getting Serious: GLM 4.7 vs. MiniMax M2.1

blog.kilo.ai2026-01-09HN 16 pts · 0 comments1,537 words

GLM 4.7 and MiniMax M2.1 successfully built a CLI task runner with 20 features in a multi-phase test. GLM 4.7 completed the task in 10 minutes with a 741-line plan and full README, whereas MiniMax M2.1 finished in 14 minutes with a 284-line plan and no documentation. MiniMax M2.1 cost $0.15 for the task and is currently free to use in Kilo Code for a limited time.written, reviewed, cited

MiniMax M2.1 used batch-based execution where tasks are grouped by dependency level.
GLM 4.7 implemented argument parsing manually to avoid external dependencies.
MiniMax M2.1 uses SHA256 for hashing, which is technically more collision-resistant than GLM 4.7's MD5.

MiniMax M2.1 Bets That ‘Most Usable’ Beats ‘Most Massive’

hackernoon.com

MiniMax-M2.1: An Open-Source Leap Toward Enterprise-Grade Agentic AI - Vanita.ai

vanitaai.com2026-02-18940 words

MiniMax M2.1: Scale and efficiency for multi-language development

tessl.io2026-01-051,329 words

MiniMax M2.1: Architecture, Benchmarks, and Practical Deployment | AiCybr Blog

aicybr.com2026-01-042,680 words

MiniMax M2.1: Open-Source MoE Model Sets Coding Benchmark - AI CERTs News

aicerts.ai2026-01-02

An Analytical Review of MiniMax M2.1 | by Barnacle Goose | Medium

medium.com2025-12-23

All coverage tracked for this model (12) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
MiniMax M2.1 Bets That ‘Most Usable’ Beats ‘Most Massive’	hackernoon.com	—	2026-10-07	exa
MiniMax-M2.1: An Open-Source Leap Toward Enterprise-Grade Agentic AI - Vanita.ai	vanitaai.com	2026-02-18	2026-10-07	exa
MiniMax M2.1: Post-Training Experience and Insights for Agent Models - MiniMax News \| MiniMax	minimax.io official	2026-01-22	2026-10-07	exa
Open-Weight Models Are Getting Serious: GLM 4.7 vs. MiniMax M2.1	blog.kilo.ai	2026-01-09	2026-10-07	hn
Training MiniMax-M2.1	minimaxi.com official	2026-01-05	2026-10-07	hn
MiniMax M2.1: Scale and efficiency for multi-language development	tessl.io	2026-01-05	2026-10-07	exa
MiniMax M2.1: Architecture, Benchmarks, and Practical Deployment \| AiCybr Blog	aicybr.com	2026-01-04	2026-10-07	exa
MiniMax M2.1: Open-Source MoE Model Sets Coding Benchmark - AI CERTs News	aicerts.ai	2026-01-02	2026-10-07	exa
MiniMax M2.1 Review: Fast, Cheap Coding AI That Actually Ships	quickleap.hashnode.dev	2025-12-29	2026-10-07	exa
MiniMax M2.1: Built for Real-World Complex Tasks, Multi-Language Programming	minimaxi.com official	2025-12-23	2026-10-07	hn
An Analytical Review of MiniMax M2.1 \| by Barnacle Goose \| Medium	medium.com	2025-12-23	2026-10-07	exa
MiniMax M2.1 Guide: Digital Employee for AI Coding	digitalapplied.com	2025-12-23	2026-10-07	exa

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

WebExplorer: Explore and Evolve for Training Long-Horizon Web Agents cited in card

Junteng Liu, Yunji Li, Chi Zhang +12 · 2025-09-08 · PDF · HF paper page (83 upvotes)

Similar models

3 models

Similar models

MiniMax-M2.5 · 229B
same size band (XXL), also MoE, same task (text-generation), more providers
MiniMax-M2.7 · 229B
same size band (XXL), also MoE, same task (text-generation), more providers
Qwen3-235B-A22B · 235B
same size band (XXL), also MoE, same task (text-generation), more providers, permissive license

Sources

4 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:35 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:02 UTC · Last verified 2026-10-25

Follow MiniMax-M2.1

Watch this model

/feeds/models/MiniMaxAI--MiniMax-M2.1.xml(RSS 2.0) · every change to MiniMax-M2.1
/feeds/families/MiniMaxAI--MiniMax-M2.1.xml(RSS 2.0) · every change across the MiniMax-M2.1 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.