MiniMax shipped M3 on June 12, 2026, making it the first open-weight AI model to offer a native 1 million token context window. The release appeared on the MiniMax homepage and Hugging Face, with weights available for download under Apache 2.0. MiniMax built M3 as a Mixture-of-Experts model with 600 billion total parameters and 32 billion active parameters per forward pass. The model is not an API-only product. Developers can run it on their own hardware or through vLLM, SGLang, and Hugging Face transformers. For teams tired of rising closed model fees, this launch follows a larger free AI model shift.
The 1M context window is the headline, but the underlying architecture matters more. MiniMax M3 uses a hybrid attention design that splits short-range and long-range context to keep inference costs stable. Total parameters sit at 600 billion with only 32 billion active, so a single 4x H100 or 8x A100 node can serve the model at acceptable speed. The context window accepts roughly 750,000 English words, enough for a full code repository, multi-day agent transcripts, or thousands of pages of legal discovery. Closed rivals like Google Gemini 2.5 Pro and Anthropic Claude now push long context through paid API tiers, but major AI model tier changes have forced developers to weigh self-hosting costs against subscription spend.
On public benchmarks, MiniMax M3 reports 85.6% on MMLU-Pro, 72.4% on GPQA Diamond, and 96.1% on MATH. Long-context retrieval quality holds up well: MiniMax claims 98.9% on RULER with 1M-token inputs and 91.2% on LongBench v2. Those numbers put M3 within a few points of closed frontier models at a fraction of the API cost once hardware is amortized. The open weights also mean no per-token price changes, no API free tier limits, and no forced migration when a vendor changes its pricing model. That is the real threat to closed providers.
Developers can pull M3 from Hugging Face today in BF16 and FP8 formats. A 4-bit GPTQ checkpoint lands the same week, which should fit on a 4x 80GB node without too much quality loss. MiniMax also published an inference stack on GitHub with vLLM and SGLang examples. Early self-hosters report 12 to 18 tokens per second on 8x A100 for 1M-token prompts, though long-context prefill remains heavy. The release lands during a broader AI price war among Google, OpenAI, and Anthropic, so open-weight 1M context may reset what developers expect from free tools.
How Do the Top Options Compare?
| Model | Context Window | Total Parameters | Active Parameters | License | Key Strength |
|---|---|---|---|---|---|
| MiniMax M3 | 1,000,000 tokens | 600B | 32B | Apache 2.0 | Long-context code and document analysis |
| DeepSeek V4 | 131,072 tokens | 671B | 37B | MIT | Math, code, low-cost inference |
| Kimi K2 | 262,144 tokens | 1,040B | 32B | Modified Apache 2.0 | Agentic coding, web tool use |
| Qwen 3.5 Max | 262,144 tokens | 720B | 36B | Apache 2.0 | Multilingual benchmarks and retrieval |
Specs reflect vendor releases and model cards as of June 2026. Long-context scores are self-reported unless benchmarked independently. Apache 2.0 in this table means commercial use allowed without royalty, but some vendors add acceptable use policies.
1. MiniMax M3 (1M Context) , Long-context research, codebase analysis, and cost-sensitive self-hosters
MiniMax M3 is the first open-weight model to ship with a 1 million token context window. The model uses a sparse Mixture-of-Experts design with 600 billion total parameters and 32 billion active parameters. That means each token only activates a small subset of experts, which keeps latency and memory use manageable for a model this size. The MiniMax homepage and Hugging Face host the BF16 and FP8 checkpoints. Developers do not need an API key to download weights. For long context, M3 splits attention into short-range sliding windows and long-range global tokens. The result is less quadratic memory blow-up. MiniMax reports 98.9% on RULER with 1M tokens, which is a strong sign that the model does not just accept long prompts but actually retrieves from them. Real-world tests on full GitHub repos and 1,500-page legal PDFs show useful retrieval, though middle-context accuracy drops slightly compared to the first 100K tokens. Self-hosting M3 is not free in practice. A 4x H100 node is the realistic minimum for decent speed. But once the GPU cost is covered, there are no per-token fees and no AI free tier limits to worry about. For a startup processing millions of long-document tokens a month, that math often beats closed API pricing.
Key strengths:
- ✅ 1M token native context window without API restrictions
- ✅ Apache 2.0 license allows commercial self-hosting and fine-tuning
- ✅ Only 32B active parameters keep inference cost lower than dense equivalents
- ✅ Strong RULER and LongBench v2 long-context retrieval scores
- ✅ BF16, FP8, and 4-bit GPTQ checkpoints available day one
- ❌ Requires 4x H100 or 8x A100 hardware for smooth long-context serving
- ❌ Middle-context retrieval quality drops slightly on very long prompts
- ❌ No hosted API from MiniMax at launch, so users manage their own inference
Who it’s for: Choose MiniMax M3 if you need a self-hosted long-context model and want to avoid per-token API pricing.
2. DeepSeek V4 , Math and code reasoning on modest open-weight hardware
DeepSeek V4 is an open-weight sparse model with 671B total and 37B active parameters. It is not a 1M context model. Its 131,072 token window covers most coding and research tasks but falls short for repository-scale analysis. The model is available under MIT license from DeepSeek. DeepSeek V4 leads on math and coding benchmarks among open models. It scores 88.1% on MATH and 80.2% on LiveCodeBench, often matching paid frontier models. For developers who care more about solving hard algorithmic problems than long-document retrieval, V4 is a better value than M3. But the smaller context window forces chunking on long codebases, which can lose cross-file dependencies. The V4 release also benefits from a huge open-source inference ecosystem. vLLM, SGLang, llama.cpp, and community quantizations were available within days. That is a meaningful advantage over MiniMax M3’s newer stack. If you want a safe, proven open-weight deployment, V4 is the lower-risk option. See best free AI models in 2026 for other no-cost options.
Key strengths:
- ✅ MIT license is among the most permissive open-weight terms
- ✅ Proven ecosystem with vLLM, SGLang, and llama.cpp support
- ✅ State of the art math and code benchmark scores
- ✅ Low active parameter count reduces serving cost
- ❌ 131K token context requires chunking for long documents
- ❌ No native 1M context retrieval
- ❌ Lags M3 on long-context RULER and LongBench v2
Who it’s for: Choose DeepSeek V4 if you need top math and coding quality and can work with chunked context.
3. Kimi K2 , Agentic coding and tool use with moderate long context
Kimi K2 from Moonshot AI is a 1,040B parameter sparse model with 32B active parameters and a 262,144 token context window. It is open-weight under a modified Apache 2.0 license. Kimi K2 focuses on agentic tool use, so it performs well when calling APIs, browsing the web, and editing code across many files. Kimi K2 holds strong scores on SWE-bench Verified and WebArena. It is designed for agents that must stay coherent over long, multi-step workflows. The 262K context is enough for many agent logs but not for 1M-token repository-scale prompts. Developers who need long-context agent traces often pair K2 with a vector database or summarization step, which adds complexity. Licensing is slightly less clean than Apache 2.0. Moonshot’s modified terms restrict some large-scale commercial use, so check the model card before deploying. This matters if you are building a paid SaaS product. For more context on agent billing and access, read Anthropic agent billing split.
Key strengths:
- ✅ Excellent agentic coding and web tool use benchmarks
- ✅ 262K context handles most multi-step agent workflows
- ✅ Strong SWE-bench Verified results
- ✅ Open weights allow local tool-calling fine-tunes
- ❌ Modified license has some commercial restrictions
- ❌ Half the context length of MiniMax M3
- ❌ Less optimized for very long document retrieval
Who it’s for: Choose Kimi K2 if your primary use case is agentic coding with tool calls, not giant document processing.
4. Qwen 3.5 Max , Multilingual tasks and low-cost retrieval workflows
Qwen 3.5 Max from Alibaba is a 720B parameter sparse model with 36B active parameters and a 262,144 token context window. It is open-weight under Apache 2.0. The model is strong across multilingual benchmarks, especially for Chinese, Arabic, and European languages. Qwen 3.5 Max excels at retrieval-augmented generation. It holds high scores on multilingual MMLU and NaturalQuestions across 20 languages. The context window is not as large as MiniMax M3, but Qwen’s efficient attention implementation makes it cheaper to serve at 262K tokens than many rivals. This is a solid option if you need broad language support without a 1M-token price tag. However, Qwen has not shipped a native 1M context version in this generation. Teams that need full-corpus reasoning will still prefer MiniMax M3. For those tracking free tier and pricing shifts across open models, see free AI pricing changes June 2026.
Key strengths:
- ✅ Strong multilingual benchmark performance
- ✅ Apache 2.0 license with no complex restrictions
- ✅ Lower serving cost at 262K context than larger-context rivals
- ✅ Large ecosystem and quantization options
- ❌ No 1M context option
- ❌ Tool use and agentic coding trail Kimi K2
- ❌ Long-context retrieval not as strong as MiniMax M3
Who it’s for: Choose Qwen 3.5 Max for multilingual RAG and lower-cost 262K context serving.
Frequently Asked Questions
Is MiniMax M3 really the first open-weight model with 1M context?
Yes, as of June 12, 2026, MiniMax M3 is the first open-weight release to ship with a native 1 million token context window. Earlier open models like DeepSeek V4 and Kimi K2 max out at 131K or 262K tokens. Some research previews have claimed long context, but none shipped open weights at this scale before M3.
What license is MiniMax M3 released under?
MiniMax M3 is released under Apache 2.0. That means you can use, modify, fine-tune, and commercially deploy the model without paying royalties. You still need to comply with standard Apache 2.0 attribution and patent terms. The weights are available on Hugging Face.
How much VRAM do I need to run MiniMax M3?
A realistic minimum is four 80GB GPUs for BF16 or FP8 inference, roughly 320GB of VRAM. A 4-bit GPTQ version cuts that to about 160GB, which can fit on two 80GB nodes or four 48GB cards. Long 1M-token prompts will use additional memory for KV cache, so plan for headroom.
Does the 1M context window actually work, or does quality drop?
MiniMax reports 98.9% on RULER at 1M tokens and 91.2% on LongBench v2. Independent early tests show strong retrieval in the first and final segments, with slight degradation in the middle. It is much better than naive long-context extensions but not perfect.
How does MiniMax M3 compare to Google Gemini 2.5 Pro or Anthropic Claude for long context?
M3 matches or beats some closed models on RULER and LongBench v2 while being self-hostable. Closed models still win on general assistant polish and instruction following. The main advantage is that M3 has no per-token API fee after hardware costs.
Where can I download MiniMax M3?
The official checkpoints are on Hugging Face under the MiniMax organization and linked from the MiniMax homepage. You can also find inference examples on GitHub for vLLM and SGLang. Download the FP8 version for the best balance of quality and VRAM use.
What Should You Remember?
- 1M context window: MiniMax M3 is the first open-weight model to process a million tokens natively.
- Apache 2.0 license: Commercial self-hosting and fine-tuning require no royalty payments.
- 600B total, 32B active: Sparse MoE design keeps serving cost lower than dense equivalents.
- 98.9% RULER: Long-context retrieval benchmarks place M3 near closed frontier models.
- 4x H100 minimum: Realistic self-hosting starts around 320GB VRAM for BF16 or FP8.
- No API fee: After hardware costs, teams avoid per-token pricing and free tier limits.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
