Shanghai-based MiniMax open sources MiniMax-M1, a model for complicated productivity tasks that supports 1M input tokens and it says beats DeepSeek's R1-0528
Chinese AI upstart MiniMax released a new large language model, joining a slew of domestic peers inspired to surpass DeepSeek in the field of reasoning AI.
Bloomberg
Context & Ripple Effects
MiniMax-M1 marks an early open-model move in MiniMax’s effort to compete in reasoning AI, with a long-context productivity focus and a stated comparison to DeepSeek. The company later continued the open-model line with an open-source M2.1 upgrade focused on coding before introducing subsequent M-series models.
The arc also shows that MiniMax’s release strategy did not remain purely open: its later proprietary M2.7 self-evolving model suggests it has used different access models as its capabilities and product ambitions developed.
First-order effects
Developers can inspect, test, and build around MiniMax-M1 for long-input productivity workloads, rather than relying only on closed reasoning-model APIs.
MiniMax directly raises the competitive benchmark it sets for DeepSeek R1-0528, while its performance claim remains the company’s own assertion.
Second-order effects
Competing Chinese model developers face added pressure to publish comparable reasoning and context-window results, and to choose whether openness is needed to win developer adoption.
Organizations evaluating models for document-heavy or multi-step work gain another candidate, increasing leverage to compare capability, deployment fit, and cost rather than standardizing on one provider.
Third-order effects
If capable reasoning models continue to be released openly, base-model capability is likely to become less defensible on its own; differentiation shifts toward distribution, tooling, reliability, and task-level economics.
MiniMax’s later mix of open releases and a proprietary M2.7 release points to a hybrid market structure in which vendors may use openness for adoption while reserving some advances for controlled products.
The trend: The release is part of a broader shift in Chinese AI toward rapid model iteration, open-model distribution, and competition on useful reasoning performance rather than model access alone.
MiniMax launch their M1 LLM as open source: - Handles 1M input and 80k output, outpacing top rivals at a fraction of the cost. - Trained using efficient methods, costing just $534,700. - Supports interactive apps with animated particle backgrounds and real-time tracking, per the …
🚨 Model Announcement: I am super excited for this new open-source LLM (MiniMax-M1), a MoE model with 1M context input window. With performance rivaling DeepSeek R1-0528 and even competing with Claude Opus 4 in some benchmarks, this is an amazing new model!! I'm eager to try it. […
Ok for sure mid-training was going to grow but I'm still having second thoughts about this. You're sure you didn't want to pretrain (??) [Minimax-M1 paper] [image]
3️⃣ Visualizations Prompt: Create an HTML page with a canvas-based animated particle background. The particles should move smoothly and connect when close. Add a central heading text over the canvas Canvas+JS, and the visuals slap.👇 [video]
This is wild. MiniMax-M1 just dropped. This AI agent = Manus + Deep Research + Computer Use + Lovable in one. 1M token memory, open weights🤯 10 wild examples + prompts & demo: 1. Netflix clone with playable trailers [video]
Compare the newly released MiiniMax-M1 80B against Claude Opus 4. MiniMax outperforms Claude Opus 4 in some benchmarks, especially in long-context driven benchmarks. [image]