MiniMax M3 Preview
MiniMax's open-weights 428B MoE with native vision, video, and sparse attention for long-context coding and agentic work.
Get API key- Provider
- MiniMax
- Price
- $0.30 in · $1.20 out / 1M
- Context window
- 524K tokens
- Released
- June 1, 2026
- License
- Open weights — available on Hugging Face
What is MiniMax M3 Preview?
MiniMax M3 Preview is an open-weights, natively multimodal large language model from Chinese AI startup MiniMax, released June 1, 2026. It uses a Mixture-of-Experts architecture with MiniMax Sparse Attention, supports vision, video, reasoning, tool use, and code generation, and offers a 524K context window on Venice.
Use it privately on Venice
On Venice, MiniMax M3 Preview runs under a private, zero-retention tier — your prompts and outputs are never stored or used for training. You can tap its native multimodal reasoning, tool use, and code-optimized capabilities without Big-Tech surveillance, while benefiting from open-weight sovereignty and permissionless access.
What can it do?
- •Native multimodal design supporting vision, video input, and multiple image inputs alongside text for deep cross-modal reasoning.
- •MiniMax Sparse Attention (MSA) enables efficient long-context processing with significantly reduced compute and memory footprint versus dense attention.
- •Strong reasoning, code-optimized generation, and built-in tool use with web search, making it well-suited for agentic coding and research workflows.
- •Open weights hosted on Hugging Face allow self-hosting, inspection, and fine-tuning outside proprietary APIs.
- •Large sparse MoE architecture (~428B parameters, ~23B activated) delivers high capacity while keeping active compute costs moderate.
- •The Venice-hosted preview offers a 524K context window — large, but shorter than the 1M token span referenced for the full M3 release.
- •As an aligned model, it is not uncensored and may refuse certain sensitive or disallowed prompts.
- •Efficient local inference requires MoE-compatible frameworks such as vLLM or SGLang; naive deployment is memory-intensive.
- •Reasoning and long-context generation can increase latency compared to smaller, non-reasoning models.
MiniMax M3 Preview capabilities
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
How to use it via API
Venice exposes an OpenAI-compatible API. Swap your base URL and call minimax-m3-preview.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-m3-preview",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Specifications
Pricing
Billed per token on Venice: $0.30 per 1M input tokens and $1.20 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
MiniMax M3 Preview vs alternatives
| Model | Context window | Open weights | Price (Venice) | Strongest at |
|---|---|---|---|---|
| MiniMax M3 Preview | 524K tokens | Yes | $0.30 in · $1.20 out / 1M | Multimodal coding & agentic reasoning |
| DeepSeek V3.2 | 160K tokens | Yes | $0.33 in · $0.48 out / 1M | General reasoning & cost efficiency |
| Kimi K2.6 | 256K tokens | Yes | $0.75 in · $3.50 out / 1M | Long-context document analysis |
| Claude Opus 4.8 | 1M tokens | No | $6 in · $30 out / 1M | Closed-source frontier performance |
The only open-weight model here with native video input, MSA sparse attention, and a 524K context window under $1.50 per million output tokens.
What is it good for?
- •Multimodal coding assistants that ingest screenshots, diagrams, and video clips alongside code for debugging and review.
- •Long-document analysis and agentic research workflows leveraging tool use, web search, and multi-step reasoning.
- •Automated UI testing and frontend development from visual mockups or screen recordings.
- •Enterprise AI workflows requiring open-weight sovereignty and private, zero-retention inference.
- •Complex knowledge-work tasks such as financial analysis or scientific literature synthesis across mixed media.
Prompting tips
- •Use the thinking parameter (enabled, adaptive, or disabled) to control reasoning depth and balance quality against latency.
- •For coding tasks, provide clear file context and specify the target language or framework to activate its code-optimized mode.
- •When using vision, upload high-resolution images or video directly; the model fuses visual and textual semantics natively.
- •For long-context prompts, place critical instructions near the beginning and end to maximize MSA attention coverage.
- •Enable function calling explicitly in your API payload when tool use is required, as adaptive reasoning may otherwise skip it.
Version history
CurrentInitial open-weights release with MSA, native multimodality, and a 524K context window on Venice.
Frequently asked questions
MiniMax M3 Preview is an open-weights, natively multimodal large language model from MiniMax, released June 1, 2026. It uses a Mixture-of-Experts architecture with MiniMax Sparse Attention, supports vision, video, reasoning, tool use, and code generation, and offers a 524K context window on Venice.
Venice bills $0.30 per million input tokens and $1.20 per million output tokens. Cached input is charged at $0.06 per million tokens. There is no subscription required; you pay per token.
The model weights are freely available on Hugging Face and can be downloaded for local deployment. On Venice you pay per token for hosted inference; there is no free tier beyond any account welcome credits.
Yes. It supports function calling, vision, video input, reasoning, web search, and code-optimized generation. You can invoke external tools via the Venice API using standard function-calling schemas.
Choose M3 Preview if you need native video input, a 524K context window, or strong multimodal coding. Choose DeepSeek V3.2 if you want lower output pricing and do not need vision or video capabilities.
MSA is a sparse attention mechanism that reduces compute and memory for long sequences by indexing key-value blocks and running full attention only on selected blocks. This cuts per-token compute compared with standard dense attention.
Yes. Because the weights are open, you can serve the model locally with vLLM, SGLang, Transformers, or KTransformers. The ~428B parameter MoE architecture requires substantial GPU memory and a framework that supports sparse routing.
No. The model carries safety alignments and may refuse certain sensitive or harmful prompts. Venice runs it privately with zero retention, but the model itself is not uncensored.
Related models
Run MiniMax M3 Preview privately.
No prompt logging. No data used for training. Free to start — no credit card.
