On 24 April, DeepSeek released the V4 preview. Two open-source models under MIT licence, 1 million tokens of context, and benchmarks that go toe-to-toe with GPT-5.4 and Claude Opus 4.6. At Delbion we are already evaluating it for our internal agents, and we have some clear takeaways.
This is not a press review. It is what matters to us as a company that implements AI agents: what it can do, what it cannot, how much it really costs, and whether it is worth it for an SME.
1. Why DeepSeek V4 matters right now
DeepSeek already surprised us with V3 in December 2024: a model trained for $5.6M that competed with GPT-4, trained for over $100M. V4 takes another step in three specific directions:
- 1 million tokens of context. That is roughly 750,000 words. An agent can read an entire contract, a full customer history or a complete codebase without fragmenting anything.
- Open-source with MIT licence. Download it, modify it, run it on your own servers. No vendor lock-in.
- Very low inference cost. V4-Pro uses only 27% of the FLOPs of V3.2 to process 1M tokens of context. The KV cache is reduced by 90%.
The combination of all three is what makes it interesting for SMEs. It is not just that it is powerful. It is that you can run it on your own infrastructure without paying OpenAI or Anthropic rates.
2. The two models: Pro and Flash
DeepSeek V4 comes in two flavours:
| Model | Total parameters | Active parameters | Best for |
|---|---|---|---|
| V4-Pro | 1.6 trillion (1.6T) | 49 billion (49B) | Complex reasoning, autonomous agents |
| V4-Flash | 284 billion (284B) | 13 billion (13B) | Fast responses, high volume, minimum cost |
Both use Mixture-of-Experts (MoE) architecture. The key difference: of the 1.6T parameters in V4-Pro, only 49B activate per token. It is like having a huge team where only the relevant specialists work on each specific task.
They also support three reasoning modes:
- Non-think: responds quickly, no deliberation. Good for routine tasks.
- Think High: analyses before responding. Slower but more accurate. The mode we use for agents.
- Think Max: maximum reasoning. For problems where quality matters more than speed.
For AI agents, Think High is the right balance
Non-think is too shallow for business decisions. Think Max burns too many resources for most operational tasks. Think High delivers quality results without excessive costs.
3. 1 million tokens: what changes in practice
The expanded context is the improvement with the most impact on AI agents. Here are concrete cases:
Customer support. An agent that previously only saw the last 20 messages from a customer can now access the full history. It detects patterns, understands the context of previous complaints, and gives responses consistent with what happened months ago.
Technical documentation. A support agent can read a complete manual (hundreds of pages) and answer specific questions without anyone having chunked it beforehand.
Software development. With 1M tokens, an agent like Claude Code or Codex can hold an entire medium-sized codebase in context. It does not need to guess which file to touch: it knows.
DeepSeek V4 achieves this with a hybrid attention architecture (CSA + HCA). It is not magic: it is engineering. The result is that processing 1M tokens costs 73% less than with V3.2.
4. The numbers: how it compares to GPT-5.4 and Claude
DeepSeek's official benchmarks place V4-Pro as the best open-source model available. Here are the figures:
| Benchmark | DeepSeek V4-Pro | GPT-5.4 | Claude Opus 4.6 | Gemini 3.1 Pro |
|---|---|---|---|---|
| LiveCodeBench (coding) | 93.5 | 92.1 | 91.8 | 89.4 |
| Codeforces Rating | 3,206 | 3,178 | 2,987 | 3,102 |
| GPQA Diamond (reasoning) | 90.1 | 89.7 | 91.3 | 87.2 |
| SWE-Verified (software agents) | 80.6 | 79.2 | 80.6 | 80.6 |
| BrowseComp (web browsing) | 83.4 | 81.7 | 82.1 | 79.8 |
| MCPAtlas Public (tool use) | 73.6 | 70.2 | 71.8 | 68.4 |
Note on these numbers
DeepSeek's benchmarks compare against maximum-effort variants of each model (GPT-5.4 xHigh, Claude Opus 4.6 Max, Gemini 3.1 Pro High). In normal use, the differences shrink. A benchmark is not a real use case.
Where V4-Pro clearly wins is MCPAtlas Public (tool use). That matters because AI agents spend most of their time calling APIs, browsing websites and executing actions, not generating pretty text.
5. What changes for AI agents
At Delbion we use agents every day. Here is what V4 allows us to do that we could not before (or was too expensive):
1. Cheaper agents without losing quality. A customer service agent using GPT-4o as a backend costs around €0.50 per conversation. With V4-Flash, that drops to under €0.05. For a company with 1,000 monthly interactions, the difference is €450 per month.
2. Memory without extra infrastructure. With 1M tokens of context, most tasks do not need an external RAG (vector memory) system. The agent keeps all relevant information within the conversation itself. Fewer moving parts, fewer things to break.
3. Local execution for sensitive data. This is critical for our clients in healthcare and finance. With an MIT licence, they can run V4-Pro on their own servers. Patient or customer data never leaves the company's infrastructure.
4. Less dependency on one provider. If DeepSeek raises prices tomorrow or changes the licence, you can migrate to another open-source model without rewriting your stack. With OpenAI or Anthropic, that vendor lock-in is a real risk.
6. What it means for an SME
If you are thinking about implementing AI agents in your company, V4 changes several important things:
Cost is no longer the barrier. V4-Flash has only 13B active parameters. It runs on accessible hardware. The DeepSeek API costs a fraction of OpenAI's. If your excuse for not automating was the price, it no longer holds.
You do not have to send your data anywhere. For a dental clinic, a tax advisory firm or an insurance broker, this matters. V4-Pro runs on your servers. GDPR-compliant without extra configuration.
Agents are good enough now. An SWE-Verified of 80.6 means an agent can solve complex software tasks. A BrowseComp of 83.4 means it can browse websites and extract information. This is not science fiction. It is what tools like Claude Code or Codex already do.
The barrier is now knowing what to automate. The model exists, it is cheap and it is good. What most SMEs lack is knowing which processes to automate, how to integrate the agent into the existing workflow, and how to measure whether it actually adds value.
Delbion data point
In the AI audits we conduct, the problem is never technological. 90% of companies have the data and infrastructure they need. What they lack is a plan: what to automate first, how to do it and who operates it. That is why we offer FUNDAE-subsidised training specifically on AI agents.
7. Costs and self-hosting
Some concrete numbers:
- V4-Pro (49B active): needs at least one A100 GPU (80GB) for smooth inference. On cloud (AWS, GCP), that is around $2-3/hour. For intensive use, buying the hardware is more cost-effective.
- V4-Flash (13B active): runs on high-end consumer configurations. An RTX 4090 (24GB) can handle it for moderate use.
- DeepSeek API: published rates are between 10x and 50x cheaper than OpenAI for the same token volume.
The quick calculation: if you currently pay €500/month on OpenAI API for your agents, with DeepSeek V4 you would pay between €10 and €50 for the same volume. The difference is not marginal. It is an order of magnitude.
8. Verdict
DeepSeek V4 is not just another model. It is proof that frontier AI can be open, cheap and accessible to companies of any size. The benchmarks are solid, the MIT licence eliminates vendor lock-in, and the inference cost makes AI agents economically viable for SMEs.
The bottleneck is no longer technology. It is organisation. Knowing what to automate, how to integrate it, and how to train your team to get the most out of it.
If you want to explore how AI agents can transform your business, we offer a free audit of automatable processes. No strings attached.
Your team needs secure AI training
The EU AI Act requires AI literacy for all staff from August 2026. Our courses cover compliance, AI agents and governance. FUNDAE can subsidise 100% of the cost.
