Questions? hello@viberation.devGet supportBlogDocsChangelog
Get started

MiniMax

Open-weight multimodal models built for long agent runs.

MiniMax publishes open weights for its M-series models, which are aimed at long-horizon agentic work and coding. The current M3 takes text, images and video with a million-token context, and the older M2 models are still around and cheaper. A solid middle option: more capable than the tiny Flash models, far cheaper than a frontier API, and you can host it yourself if you need to.

At a glance — MiniMax M3

Cost
$0.30 in · $1.20 out per 1M tokensBudget
Memory
1M tokens — about 1,550 pages
Reads
Text · Images · Video
Can do
  • Use tools (agents)
  • Structured output
  • Thinking mode
Good at
Coding: better than 66% · Agentic work: better than 67% · Overall: better than 54%
Reliability
99.96% uptime, last 24h
Released
May 31, 2026
Advanced details
API model ID
minimax/minimax-m3
Max output
512K tokens
Max input
524K tokens
Cached input
$0.06 per 1M
Tokenizer
Other
Coding index
58.6
Agentic work index
29.5
Overall index
29.2

Providers

The same model, hosted by different companies. OpenRouter picks one per request and falls back to the next if it fails.

ProviderInput /1MOutput /1MCached /1MMax outputUptime 24h
CoreWeave · fp4$0.23$0.96$0.05236K97.82%
GMICloud · fp8$0.24$0.96$0.048524K99.82%
DeepInfra · fp8$0.28$1.10$0.056512K99.83%
StreamLake · fp8$0.30$1.20$0.06512K99.57%
Venice · fp8$0.30$1.20$0.0666K93.43%
Together$0.30$1.20$0.06472K99.96%
Parasail · fp8$0.30$1.20$0.06524K99.80%
AtlasCloud · fp8$0.30$1.20$0.06524K97.74%
Novita · fp8$0.30$1.20$0.06131K99.90%
Minimax · fp8$0.30$1.20$0.06512K99.84%
MaraDegraded$0.60$2.40—944K92.87%
SambaNova$0.60$2.40—944K97.72%
ModelRun · fp4$0.75$3.00$0.15944K99.45%

Supported parameters

  • frequency_penalty
  • include_reasoning
  • logit_bias
  • logprobs
  • max_tokens
  • min_p
  • presence_penalty
  • reasoning
  • repetition_penalty
  • response_format
  • seed
  • stop
  • structured_outputs
  • temperature
  • tool_choice
  • tools
  • top_k
  • top_logprobs
  • top_p

Live from OpenRouter, refreshed hourly. Scores from Artificial Analysis.

Key info

Pricing
Open source
Category
Models
Best for
Expert
Made by
MiniMax
Weights
Open
Good for
Agentic work over a long context

More in Models

  • France's frontier lab, with the small models everyone self-hosts.

    Models
  • Meta's current line, built for long multi-agent runs.

    Models
  • The open-weight family everyone else gets benchmarked against.

    Models
  • Google's open models — Gemini's research, small enough to self-host.

    Models