DeepSeek V4: China launches a 1-trillion-parameter model in open-weight
On March 3, 2026, the Chinese startup DeepSeek — backed by the quantitative fund High-Flyer — released DeepSeek V4, a language model with 1 trillion parameters in total, of which 32 billion are active per token thanks to a Mixture-of-Experts (MoE) architecture. The model is distributed as open-weight, allowing any company to self-host it. On public benchmarks, DeepSeek V4 competes directly with GPT-5.2 and Gemini 3.1 Pro, confirming the meteoric rise of Chinese AI after the breakthroughs of DeepSeek V3 (late 2024) and DeepSeek R1 (January 2025).
What it changes for you
✦ The opportunity
DeepSeek V4 reshuffles the deck of enterprise AI. For the first time, a model rivaling the best proprietary models is available for free in open-weight. European SMBs and mid-market companies can now consider access to cutting-edge AI without depending exclusively on American APIs — and at a fraction of the cost.
Massive reduction in AI costs
By self-hosting DeepSeek V4, the cost per request drops by 80 to 90% compared to proprietary APIs. For an SMB handling 100,000 requests per month, this represents savings of €2,000 to €5,000 monthly. The MoE architecture (32 billion active parameters) ensures high performance with reasonable GPU consumption.
European technological sovereignty
With a self-hostable open-weight model, European businesses have a credible alternative to American solutions (OpenAI, Google, Anthropic). Hosted with a European cloud provider such as OVH or Scaleway, DeepSeek V4 ensures your data never leaves European soil — a decisive argument for regulated sectors.
Frontier-level performance
DeepSeek V4 matches or exceeds GPT-5.2 on the MATH-500, HumanEval and MMLU-Pro benchmarks. In multi-step reasoning, it benefits from the legacy of DeepSeek R1. Companies that used mid-tier models to save money can now access the top tier without budget compromise.
âš The risk
Data sovereignty and Chinese origin
Even in open-weight, the model was trained by a Chinese company on data whose exact composition is not public. Using the DeepSeek cloud API routes your data through servers in China, subject to Chinese intelligence law. For sensitive data, only self-hosting on European infrastructure offers sufficient guarantees. Also beware of potential cultural or censorship biases embedded during training.
Geopolitical and regulatory risk
Sino-American tensions could lead to usage or export restrictions on Chinese models in Europe. The European AI Act imposes transparency obligations on training data that DeepSeek only partially meets. Building a critical system on a single model — whatever its origin — remains a risky bet. Always plan for a multi-model architecture so you can switch quickly.
Our recommendation
DeepSeek V4 deserves serious evaluation, but with a methodical approach. Here is our three-step process:
Test on non-sensitive data
Start by evaluating DeepSeek V4 on a concrete but non-critical use case: summarizing public documents, generating marketing content, code assistance. Use the DeepSeek API (free or nearly free) for this testing phase. Measure quality and latency, and compare with your current solution on 50 to 100 representative requests.
Compare with your existing stack
Run the same test set on DeepSeek V4, GPT-5.2, Claude Opus 4.6 and Gemini 3.1 Pro. Evaluate not only the quality of the responses, but also the total cost of ownership: price per request, required infrastructure, prompt migration cost, team training. A rigorous comparison takes 2 to 3 days and avoids a hasty choice.
Assess the TCO of self-hosting
If DeepSeek V4's performance convinces you, precisely cost out self-hosting: GPU rental (4 × A100 at Scaleway or OVH), DevOps cost, maintenance, monitoring. Compare with the cost of a proprietary API over 12 months. Self-hosting becomes profitable from around 50,000 requests per month.
In summary
Frequently asked questions
Is DeepSeek V4 really open source?
DeepSeek V4 is distributed as open-weight: the model weights are freely accessible, but the training code and data remain proprietary. You can self-host it and use it commercially, but you cannot exactly reproduce the training. This is comparable to the Llama license: usable, but not truly “open source†in the strict OSI sense.
Can DeepSeek V4 be used with sensitive data?
If you self-host the model, your data never leaves your infrastructure — one of the great advantages of open-weight. On the other hand, if you use DeepSeek's cloud API, the data passes through servers located in China. For sensitive data, self-hosting or going through a European host is strongly recommended.
What infrastructure is needed to run DeepSeek V4?
The full 1-trillion-parameter model requires a cluster of at least 8 H100 GPUs with 640 GB of VRAM. In practice, most companies use the quantized version (INT4), which runs on 4 A100 80 GB GPUs, for a hosting cost of around €3,000 to €5,000 per month with a European cloud provider.
Is DeepSeek V4 better than GPT-5.2?
Public benchmarks show comparable performance: DeepSeek V4 edges out GPT-5.2 in mathematical reasoning and code, while GPT-5.2 retains an advantage in creative writing and following complex instructions. The real differentiator is cost: DeepSeek V4 self-hosted comes out 5 to 10 times cheaper than the GPT-5.2 API.
For technical profiles
Architecture and specifications
MoE architecture
1 trillion total parameters spread across 128 experts, of which 8 are activated per token, i.e. 32 billion active parameters per inference. 256k-token context window. Trained on roughly 15 trillion multilingual tokens. Architecture based on the innovations of DeepSeek V3 (Multi-Head Latent Attention, DeepSeekMoE) with an improved routing mechanism.
Lightweight version
Distilled version with 70 billion dense parameters, optimized for single-GPU deployment. Performance comparable to Claude Sonnet 4 on most tasks. Ideal as a production model for SMBs with a limited GPU budget. Latency of 0.6 s for 1,000 output tokens on an A100.
Pricing comparison
Quick comparison
| Criterion | DeepSeek V4 | GPT-5.2 | Gemini 3.1 Pro | Claude Opus 4.6 |
|---|---|---|---|---|
| Parameters (active) | 32 B (MoE) | Not disclosed | Not disclosed | Not disclosed |
| Reasoning (MATH-500) | 92.1% | 91.8% | 89.5% | 90.7% |
| Code (HumanEval) | 93.2% | 92.5% | 90.1% | 91.8% |
| Max context | 256k tokens | 128k tokens | 1M tokens | 200k tokens |
| Output cost ($/M) | $2.00 | $15.00 | $10.00 | $15.00 |
| Open-weight | Yes | No | No | No |