MiniMax open-sources M3.1 Flash weights with 428B total parameters
10/23/2026 — 10/23, 10:21·3 sources (all)·6 reports (all)
Key takeaways
- MiniMax released the open weights of MiniMax M3.1 Flash, a 428B-parameter model with 23B activated per token and across-the-board gains over…
Background & analysis
On October 19, 2026, MiniMax M3.1 Flash went live. According to an IT之家 report on October 23, the model handled more than 100 trillion tokens from real users within 72 hours of launch. On October 23, MiniMax announced on X that the weights for MiniMax M3.1 Flash were now public, and IT之家 reported the same day that MiniMax had open-sourced the model weights. The published specifications list 428B total parameters with 23B activated per token, and the company says the model delivers an overall capability improvement over the previous M3 generation.
On efficiency, IT之家 describes three optimizations. Full-model sparse attention replaces the first three full-attention layers with sparse attention. Quantization-aware training moves routed MoE experts to W4A4, with shared experts excluded, while the sparse attention indexer and main branch move from Q8KV8 to Q8KV4. Power-aware scheduling finely allocates power across compute cards to keep them running at high utilization.
In its X post, MiniMax reported scores of 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam, describing a generational leap in agentic tasks, coding and cybersecurity. On availability, IT之家 says chip vendors including Huawei Ascend, T-Head, Kunlunxin, Moore Threads, Biren, Hygon, MetaX, Iluvatar and Nvidia, as well as AMD, joined on day one, alongside cloud platforms such as Together AI, Fireworks, Nebius, Baseten and Modal. The public record currently stops at the weight release and these benchmark and partner announcements; no further developments have been reported.
AI-generated from 2 reports · updated 2 hours ago
Latest turnMiniMax released the M3.1 Flash weights as open source, with Together Compute as its day-one launch partner bringing the model to production on the AI Native Cloud. The company highlighted open weights and no GPU management required for builders.
Related AI tools
- Heat index
- 288
- All sources
- 3
- All reports
- 6
- First seen
- 2 hours ago
Reports on this story headlines open the original
MiniMax released the M3.1 Flash weights as open source, with Together Compute as its day-one launch partner bringing the model to production on the AI Native Cloud. The company highlighted open weights and no GPU management required for builders.
X：MiniMax (@MiniMax_AI)First-partyAI score 58Together AI announced that MiniMax M3.1 Flash is now live on its platform, calling itself a day-0 launch partner for the model. The model is built for agentic tasks, coding, computer use and cybersecurity, with a 1M context window and vision support. Users can try it now on Together AI.
X：Together AI (@togethercompute)First-partyAI score 48MiniMax announced a partnership with Inferact to advance production inference for its models, with engineering improvements contributed back to the vLLM project. Inferact has supported M3.1 Flash in production from day one, and the two companies said they look forward to building together.
X：MiniMax (@MiniMax_AI)First-partyAI score 43MiniMax said on X that NVIDIA's NeMo team added day-0 fine-tuning support for MiniMax-M3.1-Flash, so the model can be post-trained inside NeMo from launch. The post did not disclose parameter count, context length or pricing.
X：MiniMax (@MiniMax_AI)First-partyAI score 42MiniMax has released the weights of MiniMax M3.1 Flash, a 428B-parameter model with 23B activated, claiming across-the-board gains over the previous M3. Inference efficiency comes from sparse attention across the model, quantization-aware training and power-aware scheduling; the model processed over 100 trillion tokens in its first 72 hours after going live on October 19. Ascend, NVIDIA and AMD are among the chip vendors, alongside clouds such as Together AI and Fireworks, supporting it from day one.
IT之家AI score 66MiniMax announced on X that the weights of MiniMax M3.1 Flash are now public: 428B parameters total with 23B activated per token. The company describes a generational leap in agentic tasks, coding and cybersecurity, citing scores of 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam.
X：MiniMax (@MiniMax_AI)First-partyAI score 69
Other stories people are talking about
- Heat index 99MiniMax M3.1 Flash targets interruptible voice agents3 sources
- Heat index 1173PHP ships emergency patch for two critical AI-found flaws22 sources
- Heat index 901RisingMeta Creates Non-Voting Class C Shares to Fund AI Spending14 sources
- Heat index 738Anthropic launches Claude Duet agent in research preview14 sources
- Heat index 534DeepSeek's second funding round nears close at a reported $70B pre-money valuation11 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion