MiniMax M3: 428B MoE, 1M Context, Multimodal Coding Agent (2026)

Last updated July 2026.

MiniMax M3 is an open-weight mixture-of-experts model from MiniMax, released June 1, 2026. 428B total parameters, ~23B active per token, a 1M-token context window, and native image and video input. The attention stack is grouped-query attention with MiniMax Sparse Attention (MSA), which cuts per-token compute at 1M context to about 1/20 of the previous generation. It matches Claude Sonnet 4.6 on real-world agentic benchmarks. On Morph, morph-minimax3-428b runs at $0.60/M input and $2.40/M output with a 256K context.

What Is MiniMax M3?

MiniMax M3 is an open-weight mixture-of-experts (MoE) model from the Shanghai lab MiniMax, released June 1, 2026. It has 428 billion total parameters with roughly 23 billion active per token, a 1-million-token context window, and native image and video input. MiniMax positions it as the first open-weight release to combine reasoning, agentic coding, and multimodality in a single model. Weights went live on Hugging Face by June 7, and the technical report landed on arXiv on June 11.

The multimodality is not bolted on. MiniMax trained M3 on mixed text, image, and video from step 0, and the model can operate a desktop from screenshots. That is the structural difference from DeepSeek V4, GLM 5.2, and Qwen 3.6, all text-only. Morph serves it as morph-minimax3-428b, one of its fast general coding models, on custom kernels via an OpenAI-compatible API. See pricing.

MiniMax M3 Architecture: 428B MoE + MSA

MiniMax M3 is a 428B-parameter sparse MoE with about 23B parameters active per token, built on a grouped-query attention backbone with MiniMax Sparse Attention (MSA) layered on top. MSA is a new sparse-attention operator designed for million-token contexts: MiniMax reports it partitions KV into blocks more precisely than DSA or MoBA, using a "KV outer gather Q" access pattern that cuts attention compute and memory while preserving quality.

MSA speedups vs MiniMax M2

The whole point of MSA is making a 1M-token context economical to serve. At 1M context, MiniMax reports M3 runs more than 9x faster in the prefill stage and more than 15x faster in the decode stage versus MiniMax M2, with per-token compute at roughly 1/20 of the previous generation. That is what lets long context be used inside an agent loop that makes many calls, rather than being a spec you cannot afford to fill.

Benchmarks vs GLM 5.2, DeepSeek V4, GPT-5.5

MiniMax M3 scores 80.5% on SWE-bench Verified and 66.0% on Terminal-Bench 2.1, matching Claude Sonnet 4.6 on real-world agentic benchmarks. On Artificial Analysis's Intelligence Index (v4.1) it sits at 44, tied with DeepSeek V4 Pro and behind GLM 5.2 at 51. So on raw text intelligence it is a mid-pack open-weight model; its separation comes from multimodality and price, not benchmark rank.

Where M3 lands among open-weight models

The honest read from independent trackers: for text-only agentic coding, GLM 5.2 is the better default, and DeepSeek V4 Flash is cheaper still on pure cost. OpenRouter's June 2026 open-weights roundup states plainly that M3 is "not the best pure coding model in the group." Its moat is the one thing none of the others do: reading image and video natively.

Read the AA Index column against the SWE-bench column: the three big MoEs cluster within a point of each other on SWE-bench Verified (80.5 to 80.9), yet GLM 5.2 pulls 7 points ahead on the aggregate index. That gap is planning and long-horizon reasoning, not the ability to write one patch. GLM figures are secondary-source estimates; check the live Artificial Analysis board before committing volume.

The Interleaved-Thinking Tool-Calling Gotcha

MiniMax M3 uses interleaved thinking: it reasons between tool calls inside <think> tags, the pattern Anthropic introduced with Claude Sonnet 4. The footgun most people hit first is that MiniMax's own docs require you to feed the model's reasoning back on every turn. Strip the thinking content between tool calls and multi-turn tool use degrades. This is the single most common M3 integration bug, inherited straight from M2.

Practical fixes: on the Anthropic SDK path, append the full response.content list (all blocks: thinking, text, tool_use) to the message history. On the OpenAI-compatible path, echo back the assistant message with its <think> content intact rather than dropping it. If your harness has a "strip reasoning" or "drop thinking" setting, turn it off for M3. This one change is often the difference between M3 looking broken and M3 matching its benchmark scores in a real Cline or Claude Code loop.

The other real-world caveat, from OpenRouter's June 2026 roundup: M3 "can be verbose, and reasoning-heavy," so a low per-token price does not guarantee cheap operations. A reasoning-heavy model that emits long thinking traces on every turn spends output tokens fast. Budget on tokens actually generated in your workload, not on the headline output rate.

Pricing and How to Run MiniMax M3 on Morph

MiniMax's official API is tiered by input length, starting at $0.30/M input and $1.20/M output and rising above 512K-token context. OpenRouter's weighted-average realized price is about $0.098/M input and $1.21/M output. Morph serves morph-minimax3-428b at $0.60/M input and $2.40/M output with a 256,000-token context, on custom low-level kernels tuned for code generation.

Point any OpenAI-SDK client at https://api.morphllm.com/v1 and pass the model id. Because Morph is OpenAI-compatible, M3 drops into existing tooling without a proxy:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.morphllm.com/v1",
    api_key="$MORPH_API_KEY",
)

resp = client.chat.completions.create(
    model="morph-minimax3-428b",
    messages=[
        {"role": "user", "content": "Refactor this React component to use hooks."},
    ],
)
print(resp.choices[0].message.content)

For multi-turn tool calling, echo the full assistant message (including any <think> content) back into messages on the next turn, per the interleaved-thinking rule above. See Morph Open Source Models for the full lineup and pricing for every model's rate.

MiniMax M3 vs M2.7: What Changed

M3 is a larger, multimodal successor to the M2 line. M2 was a 230B MoE with about 10B active per token, purpose-built for agentic coding, and it topped open-weight agentic benchmarks cheaply on release. M3 nearly doubles total parameters to 428B, roughly doubles active params to ~23B, adds native image and video, and swaps in MSA to keep a 1M context affordable. Morph serves the predecessor as morph-minimax27-230b (MiniMax M2.7) at $0.279/M input and $1.20/M output with a 196,608-token context.

Both models share the interleaved-thinking format and the same tool-calling rule, so the "don't strip the reasoning" guidance applies to either. If you integrated M2 correctly, M3 needs no new plumbing beyond the model id, plus wiring up image and video inputs if you want the multimodal path.

When to Use MiniMax M3 (and When Not)

Use M3 when your agent needs to read screenshots, video, or UI state, or when you want a cheap 1M-context multimodal backend and can preserve reasoning between turns. Reach for GLM 5.2 when the work is text-only planning and long-horizon coding, or DeepSeek V4 Flash when raw cost per token is the deciding factor. A common pattern is routing: multimodal and UI tasks to M3, hard text reasoning to GLM 5.2.

Limitations

Frequently Asked Questions

What is MiniMax M3?

An open-weight mixture-of-experts model from MiniMax, released June 1, 2026. 428B total parameters, ~23B active per token, a 1M-token context, and native image and video input. It uses a GQA backbone with MiniMax Sparse Attention (MSA) and scores 80.5% on SWE-bench Verified. It is the first open-weight model to pair frontier agentic coding with native multimodality.

What is MiniMax M3's context window and parameter count?

428B total / ~23B active, 1M-token context officially. MSA gives it >9x faster prefill and >15x faster decode versus M2 at 1M context, at about 1/20 the per-token compute. On Morph, morph-minimax3-428b serves a 256K context.

How does MiniMax M3 compare to GLM 5.2 and DeepSeek V4?

On the AA Intelligence Index, M3 scores 44, tied with DeepSeek V4 Pro and behind GLM 5.2 (51). GLM 5.2 is the stronger text-only coding default; DeepSeek V4 Flash is cheaper on pure cost. M3's edge is native image and video, making it the pick for multimodal and UI-agent work.

Why does MiniMax M3 break in multi-turn tool calling?

Because it uses interleaved thinking and MiniMax's docs require you to append the model's full response, including <think> content, back to history every turn. Frameworks that strip reasoning between tool calls break the chain and degrade multi-turn tool use. Preserve the thinking blocks verbatim.

How much does MiniMax M3 cost?

Official API: from $0.30/M input and $1.20/M output, rising above 512K context. OpenRouter weighted-average: about $0.098/M input, $1.21/M output. Morph: morph-minimax3-428b at $0.60/M input and $2.40/M output, 256K context.

Is MiniMax M3 open source?

The weights are open on Hugging Face (MiniMaxAI/MiniMax-M3) under a custom minimax-community license the community noted was more permissive than M2.7's. The technical report is on arXiv (June 11, 2026). At 428B total, self-hosting is multi-node territory.

Can I use MiniMax M3 in Claude Code or Cline?

Yes. Point the tool at an OpenAI- or Anthropic-compatible endpoint serving M3 and select the model. The one thing to get right is preserving the model's <think> content between tool calls; disable any "strip reasoning" setting so interleaved thinking works.