DeepSeek V3.1: the silent revolution of open source arrives under the MIT license

DeepSeek V3.1: the silent revolution of open source arrives under the MIT license

🔎 A Chinese model, an American license, an earthquake in AI

May 2026. Open-source AI has just lost its last locks. DeepSeek drops V3.1, a 671-billion-parameter model under the MIT license — the most permissive in software. Zero commercial restrictions, zero sharing obligations, zero usage limits.

This is as much a geopolitical maneuver as a technical one. A Chinese company adopts Silicon Valley's favorite license to distribute a model that rivals GPT-5 and Claude 4.1 on most benchmarks. The message is clear: open source no longer has borders.

The other major novelty is hybrid reasoning. DeepSeek V3.1 decides on its own whether it needs to think deeply or answer instantly. No more need to choose between a slow, accurate model or a fast, superficial one. It does both, automatically.

The essentials

• MIT license: the most permissive on the market, allowing all commercial use without any quid pro quo. No model of this scale had ever been released under this license.
• Hybrid reasoning (Think & Non-Think): the model automatically switches between deep reasoning and fast response depending on the complexity of the query.
• 671B parameters, 37B active: thanks to the MoE (Mixture of Experts) architecture, only a fragment of the model is activated at each inference, drastically reducing costs.
• Proprietary-level performance: according to AIPure, DeepSeek V3.1 sits in the same category as GPT-5 and Claude 4.1 on standard benchmarks.
• 128K context tokens: sufficient for most professional use cases, from document analysis to complex code.

Recommended tools

What the MIT License Really Changes

An MIT license, in the AI world, is a declaration of war against proprietary models. Not in an aggressive sense — in a structural sense.

The Apache 2.0 license, used by Meta's Llama, imposes conditions: license notices, no trademarks, and a patent defense mechanism. MIT requires none of that. You take the model, modify it, sell it, integrate it into a commercial product, and you owe absolutely nothing to anyone.

This is exactly what BentoML confirms in its comprehensive guide to DeepSeek models: DeepSeek-V3 Base and Chat are open-source and commercially usable under the MIT license, without restriction.

For a startup that wants to integrate an LLM into its product without relying on an external API, it's the holy grail. You download the weights, deploy them on your servers, and you don't have to worry about a provider changing its prices overnight or cutting off your access.

Geopolitics makes it even more striking. A model developed in China, released under the most "American" license possible, available to any company on the planet. Attempts to control AI chip exports did not prevent DeepSeek from producing a world-class model. Proof, if any were needed, that regulatory barriers alone are no longer enough.

Hybrid reasoning: Think & Non-Think

This is the most significant technical innovation of this V3.1. Until now, the market had two types of models: those that think for a long time (DeepSeek-R1, OpenAI's o1) and those that answer quickly (GPT-4o, Claude Haiku).

DeepSeek V3.1 merges these two approaches into a single model. The official announcement from DeepSeek mentions "hybrid inference" with two modes: Think (deep reasoning) and Non-Think (direct response). The switch is done via templates at the API level.

In practice, when you ask a simple question ("What is the capital of France?"), the model answers in a fraction of a second without activating its reasoning chain. When you give it a complex code problem or a multi-step logical deduction, it activates its Think mode and unfolds a step-by-step reasoning.

The result reported by John Rhodes on LinkedIn: DeepSeek-V3.1-Think is faster than DeepSeek-R1-0528 while maintaining an equivalent level of reasoning. You lose less in latency, you gain in flexibility.

For developers, this changes the game. A single endpoint, a single model to maintain, and it is the system that decides the level of cognitive effort to provide. The meilleurs LLM pour coder are increasingly integrating this adaptive reasoning logic anyway.

Benchmarks: Where DeepSeek V3.1 Stands Against the Giants

The numbers speak for themselves. AIPure publishes a detailed comparison of DeepSeek V3.1 against GPT-5 and Claude 4.1. The verdict: the open-source model doesn't consistently outperform them, but it sits in the same performance zone on the majority of tests.

The benchmark table according to OpenRouter shows a 671B parameter model of which only 37B are active at each step. This MoE architecture is key: it allows the power of a giant model with the inference cost of an average model.

In the monthly comparison of the best LLMs, DeepSeek V4 Pro (Max) dominates the open-source ranking with a score of 88, followed by Kimi K2.6 at 85. But V3.1 remains relevant for deployments where inference cost is critical, as V4 Pro requires significantly more resources.

Positioning in the open-source ecosystem according to the 2026 open-source model ranking :

DeepSeek V3.1 is not aiming for the absolute top of the ranking. It aims for the best performance/cost/license ratio. And on this trio, it has no direct competitor.

Concrete impact for startups and businesses

The real question isn't "is V3.1 better than GPT-5?" but "is V3.1 sufficient for my use case, at a cost 10 to 50 times lower?". For 90% of businesses, the answer is yes.

Let's take a real-world case: a SaaS startup that wants to integrate an AI assistant into its product. With the OpenAI API, it pays per request, depends on service availability, and controls neither the data nor the model. With DeepSeek V3.1 under the MIT license, it can host the model itself or go through an alternative API provider, and retain total control.

Inference costs are radically different. Only 37 billion of the 671 billion parameters are active at each request. This means that a server with 4-8 high-end GPUs can run V3.1 reasonably, whereas a dense 671B model would be completely out of reach.

For businesses that want to go further and run des LLM en local, the guide to installation de LLM local remains relevant, even if V3.1 will require substantial hardware. Lighter meilleurs LLM locaux like Qwen3.6-27B remain better suited for modest configurations.

The impact on the AI value chain is tangible. When a model of this quality is free and unrestricted, the value no longer lies in the model itself but in what is built around it: the interface, specific training data, workflows, integration into real products.

DeepSeek V3.1 vs. Qwen, Llama, and the rest of open source

The open-source landscape in 2026 is dominated by three main pillars: DeepSeek, Alibaba (Qwen), and Meta (Llama). Each has a different strategy.

Qwen3.6 by Alibaba offers a very comprehensive family of models, from 27B to 397B, under the Apache 2.0 license. This is the "one size does not fit all" approach: you choose the model suited to your hardware. Qwen3.6-27B scores 74 on the benchmark, which is remarkable for its size.

DeepSeek takes the "one massive but optimized model" approach. V3.1 is 671B but only activates 37B. It is more powerful than Qwen3.6-27B in absolute terms, but also heavier to deploy. The MIT license also gives it a clear legal advantage over Apache 2.0 for the most aggressive commercial uses.

Meta's Llama 4, absent from the top of the 2026 open-source rankings, seems to have lost its momentum. The void left by Meta at the top has been filled by DeepSeek and Kimi.

What makes DeepSeek V3.1 particularly important is the combination of the MIT license + hybrid reasoning + high performance. No other open-source model checks these three boxes simultaneously. This is why it is probably the most strategic open-source model of the year.

For developers looking for the best free LLMs, accessing DeepSeek V3.1 via OpenRouter is a solid option, with rates well below those of equivalent proprietary models.

AI agents: why V3.1 is laying the groundwork

DeepSeek's official announcement emphasizes "enhanced agentic capabilities": tool use, multi-step tasks, the ability to chain complex actions. V3.1 is not an agent in itself; it is the engine that makes it possible to build them.

Hybrid reasoning is particularly well-suited for agents. An AI agent sometimes needs to respond instantly ("here is the requested file") and sometimes plan over multiple steps ("analyze this repository, identify the bugs, suggest fixes, generate the patches"). Having a single model that handles both regimes seamlessly greatly simplifies the architecture.

It is in this context that projects like ByteDance's DeerFlow make complete sense. DeerFlow is an open-source agent that researches, codes, and creates over the long term. Pairing it with a model like DeepSeek V3.1, capable of switching between fast and deep reflection, is a natural combo.

The trend towards AI agents with local Ollama is also accelerating. Developers want agents they control end-to-end, without cloud dependencies. V3.1 under the MIT license is an ideal candidate as the brain for these agents, even if its local deployment requires considerable hardware.

The best LLMs for AI agents are increasingly integrating this adaptive reasoning capability as a key selection criterion.

The geopolitical context: open source AI no longer has borders

To be perfectly clear: a model of this class, developed in China, released under an MIT license, and downloadable by anyone on GitHub, renders a large part of American attempts to control AI obsolete.

Export restrictions on NVIDIA chips aim to slow down Chinese AI development. DeepSeek V3.1 is proof that these measures have limited effects. The company has optimized its infrastructure to work with the available chips, and the result is a model that rivals the best American products.

The MIT license acts as a peaceful geopolitical Trojan horse. No restrictive license that would limit distribution to certain regions. No embargo clause. An open model, universally accessible, that American, European, African, or South American companies can integrate into their products without any legal friction.

For the meilleurs LLM en français, the impact is indirect but real. The more high-quality open source spreads, the more Francophone communities can fine-tune these models for local use cases without depending on Anglo-Saxon players.

Deploying DeepSeek V3.1: what you need to know

Deploying V3.1 is no trivial task. 671 billion parameters, even with only 37B active, require serious infrastructure.

Via API (recommended for most)

OpenRouter offers DeepSeek V3.1 with both thinking and non-thinking modes accessible via specific templates. This is the simplest way to use it in production. Pricing varies depending on the mode (thinking costs more in tokens because the model generates its reasoning chain).

Self-hosting

The official GitHub repo provides the weights and instructions. Plan for a minimum of 8 A100 80GB GPUs for a comfortable deployment, or the equivalent in consumer GPUs if you are willing to accept higher latency.

Quantization can significantly reduce memory requirements, but at the cost of a measurable degradation in performance on complex reasoning tasks. It is a trade-off to evaluate based on your use case.

For more modest configurations, the best local LLMs like Qwen3.6-35B-A3B (score 67 with only 3B active) remain more realistic alternatives. The arrival of Qwen3.5-122B-A10B with only 10B active parameters also shows that the optimization trend is accelerating across the entire ecosystem.

❌ Common mistakes

Mistake 1: Confusing V3.1 and V4

DeepSeek V4 Pro tops the benchmarks with a score of 88. V3.1 is lower, around 70-75. If your priority is raw performance, look at V4 first. V3.1 shines in the performance/cost/license ratio, not in the absolute score.

Mistake 2: Underestimating hardware requirements for self-hosting

671B parameters is huge. Even with 37B active, loading the weights into memory requires several hundred GB of VRAM. Do not start a local deployment without checking your resources. For local use on a standard machine, turn to lighter models.

Mistake 3: Ignoring hybrid mode and forcing thinking

The classic mistake is systematically enabling Think mode to "get the best answers". This is counterproductive: for simple questions, you add latency and cost without any quality gain. Let the model decide, or use Non-Think mode by default and only switch to Think for complex tasks.

Mistake 4: Believing that the MIT license allows everything without limits

The MIT license covers the model itself. It does not cover the training data (which may have its own restrictions), nor does it protect you from local regulations like the European AI Act. Free does not mean without responsibility.

❓ Frequently Asked Questions

Is DeepSeek V3.1 really free for commercial use?

Yes. The MIT license places no restrictions on commercial use. You can integrate V3.1 into a SaaS product, modify it, resell it, without paying any license fees or royalties to DeepSeek. This is confirmed by BentoML in their analysis of the DeepSeek ecosystem.

What is the difference between Think and Non-Think mode?

Think mode activates deep reasoning (visible chain of thought, slower but more accurate response). Non-Think mode provides a direct, fast response without any intermediate steps. The model switches between the two based on the templates sent via the API, as detailed in the official announcement.

V3.1 or V4 Pro, which one to choose?

V4 Pro for maximum performance (score of 88). V3.1 for the best performance/cost ratio, especially in self-hosting where the difference in resources between the two generations is significant. For the best LLMs for research, V4 Pro is probably more suitable thanks to its better understanding of complex queries.

Can DeepSeek V3.1 be fine-tuned?

Yes, the MIT license explicitly allows it. In practice, fine-tuning a 671B MoE model is a heavy engineering project. LoRA/QLoRA techniques lower the barrier, but remain resource-intensive. For accessible fine-tuning, smaller models from the Qwen family or DeepSeek V4 Flash are more realistic.

Does DeepSeek V3.1 handle French well?

The published benchmarks are primarily in English. Like any model of this scale trained on multilingual data, V3.1 understands and generates French satisfactorily, but is not specifically optimized for this language. For high-quality native French, fine-tuning or the use of a specialized model remains preferable.

✅ Conclusion

DeepSeek V3.1 is not the most powerful model of 2026 — that role falls to its little brother V4 Pro. But V3.1 is probably the most strategic: MIT license, hybrid reasoning, respectable performance, controlled costs. It is the model that makes AI independence accessible to companies that cannot afford to rely on proprietary APIs. The rest of the best LLM ranking can tremble.