← Models

MiniMax M3

MiniMax · MiniMax M

MiniMax's 1M-context native multimodal open-weight model for coding and long-horizon agents.

Part of MiniMax M family · Other versions: MiniMax M2.7 , MiniMax M2.5 , MiniMax-M2.5
Type
multimodal
Context
1M tokens
Max Output
N/A
Status
current
Input
$0.3/1M tok
Output
$1.2/1M tok
API Access
Yes
License
MiniMax Community License
multimodal coding agentic tool-use computer-use long-context sparse-attention open-weights image-input video-input
Released June 2026 · Updated July 10, 2026

Overview

Freshness note: Model capabilities, limits, licensing, and pricing can change quickly. This profile is a point-in-time snapshot last verified on July 10, 2026.

MiniMax M3 is MiniMax’s current flagship for coding, agentic work, and long-horizon multimodal tasks. Released on June 1, 2026, it moves the M family from proprietary text-only API models into a native multimodal generation with public weights, a 1M-token total context budget, and text, image, and video input.

The model is available through MiniMax’s API and coding products as well as official open-weight downloads. That makes it relevant both as a low-cost managed route and as a deployment-controlled model for teams with the infrastructure to run a very large mixture-of-experts checkpoint.

Capabilities

MiniMax positions M3 for software engineering, tool use, research, desktop automation, and other workflows that must keep working across many steps. Its multimodal input support is useful for visual debugging, interface inspection, document analysis, and video understanding while still returning text.

Three reasoning modes are exposed through the API’s thinking setting: enabled, adaptive, and disabled. This lets applications trade reasoning depth against latency rather than committing every request to the same execution style.

Technical Details

Official anchors at this snapshot:

  • Model ID: MiniMax-M3.
  • Approximately 428B total parameters and 23B active parameters in the official open-weight model card.
  • Up to 1,000,000 total tokens across input and output; MiniMax describes 512K as the guaranteed minimum supported context.
  • MiniMax Sparse Attention, designed to reduce long-context computation.
  • Native text, image, and video input with text output.
  • Public weights under the MiniMax Community License.

MiniMax does not publish a separate durable maximum-output limit for M3. Here, maxOutput: 0 means “not separately published,” not zero output capacity. Benchmark generation settings are not treated as production API limits.

Pricing & Access

MiniMax’s current pay-as-you-go API pricing is tiered by total input length:

  • Up to 512K input: 0.30per1Minputtokens,0.30 per 1M input tokens, 1.20 per 1M output tokens, and $0.06 per 1M cache-read tokens.
  • Above 512K through 1M: a published long-context rate of 0.60input,0.60 input, 2.40 output, and $0.12 cache read per 1M tokens.

The listed base rates apply to the standard up-to-512K tier. At this snapshot, hosted input above 512K is available only in limited quantity for a limited time and requires contacting MiniMax sales; it is not yet an ordinary self-serve tier. MiniMax also offers MiniMax Code and subscription token plans, while the official Hugging Face release supports self-managed inference.

The weights are publicly downloadable, but the MiniMax Community License is not an open-source license such as Apache 2.0 or MIT. Its default grant is for noncommercial use, while commercial use adds display and notice requirements and requires prior authorization above the license’s stated annual-revenue threshold. The license also includes use restrictions, so teams should review it before distributing or monetizing a deployment.

Best Use Cases

Choose MiniMax M3 for repository-scale coding agents, long-running research or tool loops, multimodal software work, and high-volume API workloads where the current token price is attractive. It is also a candidate for private infrastructure when a team can support its very large weight footprint.

It is less natural for casual laptop deployment, teams that need a permissive open-weight license, or applications that require a provider-published output-only cap for strict capacity planning.

Comparisons

  • MiniMax M2.7: Older text-only hosted tier with the same base token price and a smaller context; M3 adds multimodality, 1M context, and public weights.
  • GLM-5.2: MIT-licensed 1M-context coding peer; M3 adds native image and video input and a lower direct API price, while GLM has the more permissive license.
  • Qwen3.6-27B: Far smaller Apache-licensed multimodal model for more practical self-hosting; M3 targets a much higher infrastructure and capability tier.