Skip to main content
Back

Factory's Droid Now Runs MiniMax M3 at 5% the Cost of Closed Models

MiniMax M3, the first open-weight model combining frontier coding, 1M-token context, and native multimodality, is now available inside Factory's Droid agent platform

Factory just added MiniMax M3 to its Droid coding agent platform, giving developers access to one of the most capable open-weight models available today without leaving their existing workflow. The move is a signal of where agentic development tooling is heading: model-agnostic platforms that let you route tasks to the right engine rather than betting everything on a single provider.

What makes M3 different from every other open model

M3 is positioned as the first open-weight model with three frontier capabilities, achieving top-tier performance on coding and agentic benchmarks with autonomous task decomposition, tool invocation, and multi-step reasoning. That combination has historically been exclusive to closed-source labs. Until now, only a handful of closed-source models could simultaneously achieve frontier coding capabilities, million-token context, and multimodality. M3 is the first to bring that complete package to the open world.

The three pillars are:

  • Frontier coding and agentic performance -- On SWE-Bench Pro, a harder benchmark built around 1,865 real pull requests from 41 actively maintained open-source repositories, MiniMax reports M3 scored 59.0%, ahead of GPT-5.5 at 58.6% and Gemini 3.1 Pro at 54.2%.
  • 1M-token context window -- Powered by the proprietary MiniMax Sparse Attention (MSA) architecture, M3 supports up to 1M tokens, which serves as the infrastructure for long-range agent tasks, long-range coding, and long-video understanding.
  • Native multimodality -- Rather than fusing a separate vision model onto a pretrained text network, MiniMax engineered M3 as a natively multimodal system from step zero, blending interleaved sequences of text, images, and visual components across a pretraining corpus beyond 100 trillion tokens. This deep alignment enables the model to translate complex visual geometries, such as programming charts or coordinate maps, into structural code without losing contextual fidelity.

The architecture that makes a 1M context window actually usable

A million-token context window sounds impressive until you realize most models become prohibitively slow and expensive at that scale. MiniMax's answer is MSA (MiniMax Sparse Attention), a new attention mechanism that sidesteps the quadratic cost explosion of standard full attention by pre-filtering which parts of the context actually need to be attended to.

MSA replaces full attention with KV-block selection to cut per-token compute at long context, achieving roughly 1/20 the cost of the previous generation at 1M tokens, with substantially faster prefill and decode while retaining quality across most tasks. In practice, the MSA architecture cuts per-token compute at 1M context to one-twentieth of the prior generation, with more than 9x faster prefill and more than 15x faster decoding.

What it can actually do in the wild

Benchmarks are one thing. MiniMax published two internal stress tests that are worth paying attention to.

The first was a CUDA kernel optimization challenge. In the task, M3 ran for about 24 hours, made 147 benchmark submissions, used 1,959 tool calls, and kept improving after long performance plateaus. After six landmark rounds of optimization, M3 improved Hopper FP8 hardware peak utilization from 7.6% in the first version to 71.3%, achieving a 9.4x speedup compared with the original version.

The second was autonomous model training. MiniMax gave M3 four pretrain-only base models and asked it to autonomously complete the full pipeline -- data synthesis, training, evaluation, and iteration -- within 12 hours, with the entire process running without human intervention. M3 scored 37.1, ranking third overall, behind only Claude Opus 4.7 and GPT-5.5.

Beyond the headline numbers, M3 also posted:

  • 66.0% on Terminal Bench 2.1, 74.2% on MCP Atlas, and 83.5 on BrowseComp, outstripping Claude Opus 4.7's score of 79.3 in autonomous browsing and information retrieval.
  • 70.0% on OSWorld-Verified, which measures a model's ability to operate desktop graphical interfaces.

Where it falls short

The benchmark story has a real asterisk. Benchmark scores are company-reported and run on MiniMax's own infrastructure, and the promised open weights had not been released at launch. Several results were run on MiniMax's own infrastructure with agent scaffolding, so independent verification is still pending.

When contrasted with Anthropic's Claude Opus 4.8, M3's competitive ceiling becomes evident: in the domain of pure code modification on SWE-Bench Pro, M3's 59.0% score drops behind Opus 4.8's leading 69.2%. For the most demanding, high-stakes engineering work, the closed frontier still leads.

The cost equation that changes the calculus

This is where M3 becomes genuinely interesting for production use. Via OpenRouter, M3 is priced at $0.30 per million input tokens and $1.20 per million output tokens, with a 1,048,576-token context window and maximum output of 512,000 tokens. With cache optimization, the blended cost can drop as low as $0.06 per million tokens.

For context, that puts frontier-class coding performance at roughly 5-10% of what you'd pay for comparable closed models. Analyzing a 10,000-line codebase (approximately 40K input tokens, 10K output tokens) costs around $0.024.

MiniMax also offers subscription tiers for heavier usage: Plus at $20/month for roughly 1.7B tokens, Max at $50/month for ~5.1B tokens, and Ultra at $120/month for ~9.8B tokens, with text, image, speech, and music all drawing from the same pool.

How to use it inside Factory Droid

Droid is model-agnostic and can change models mid-session, with models from all major providers including Anthropic, OpenAI, Google, and NVIDIA. M3 slots into that roster as a cost-efficient open-weight option. You can switch to it mid-session with /model without losing your chat history.

For those who want to wire it up manually via BYOK (Bring Your Own Key), MiniMax exposes an Anthropic-compatible API endpoint, so the configuration in ~/.factory/config.json looks like this:

{
  "custom_models": [
    {
      "model_display_name": "MiniMax-M3