DeepSeek’s Free Models Shake Up AI Race

What happened

This week DeepSeek, the Hangzhou‑based AI startup that became a viral sensation earlier in 2025, released two new models—DeepSeek‑V3.2 and a high‑reasoning variant called DeepSeek‑V3.2‑Speciale—and made the weights and code broadly available under a permissive open‑source licence. The company positions the pair as models tuned for long documents and multi‑step problem solving; in public benchmarks and contest simulations it claims performance comparable to the newest proprietary frontier systems.

These are not small updates. DeepSeek describes them as a step change in long‑context efficiency and agentic tool use, and the company has published model cards, a technical report and downloadable weights for developers and researchers to experiment with.

How the models work — and why they cost less to run

The headline innovation DeepSeek highlights is a form of sparse attention they call DeepSeek Sparse Attention (DSA). Attention mechanisms are the part of large language models that let them weigh which words and passages matter for a given answer. Traditional attention scales poorly with input length—the compute cost grows roughly with the square of the number of tokens—so feeding thousands or tens of thousands of tokens becomes prohibitively expensive.

Benchmarks, competitions and real‑world tasks

DeepSeek has published a mix of standard benchmarks and more dramatic contest‑style evaluations. The Speciale variant is presented as a deep‑reasoning engine tuned through reinforcement learning and specialized training regimes; in the company’s reported numbers it achieves gold‑medal level performance on several elite programming and mathematics competitions and posts competitive results on coding and reasoning benchmarks that are typically used to compare frontier models.

Those contest results are striking on paper: DeepSeek’s materials report high scores on mathematics and informatics olympiad problems taken under test‑like constraints, and it shows strong performance on coding workflow benchmarks. If the numbers hold up under independent review, they indicate that a smaller set of architectural changes and targeted training can deliver reasoning gains without simply scaling compute forever.

Agentic "thinking with tools"

A second practical advance DeepSeek emphasises is preserving internal reasoning when the model interacts with external tools—search, code execution, file editing and so on. Earlier models tend to lose their internal chain of thought each time they call an external API; DeepSeek teams this with a training pipeline of synthetic multi‑step tasks so the model learns to maintain and carry forward partial plans while it queries tools. That makes multi‑step workflows—debugging complex code, planning logistics with changing constraints, or navigating research across many documents—far smoother in practice.

The training regimen DeepSeek describes includes thousands of synthetic environments and task variations intended to teach the model how to deliberate and act in tandem. For developers building autonomous agents or assistant workflows, that capability matters as much as raw benchmark scores: it reduces the engineering friction of stitching tools and models together.

Unlike most companies that keep their largest models behind paid APIs, DeepSeek has released model weights and code under an MIT‑style licence and published integration examples for popular runtimes. That move lowers the bar for deployment—enterprises can run the models on‑prem, researchers can inspect logits and failure modes, and startups can build agents without the same vendor lock‑in concerns.

The combination of open weights plus efficiency improvements matters commercially: lower inference costs and self‑hosting options change both unit economics and risk calculations for customers that need heavy use of long‑context reasoning (legal discovery, software ingest, scientific literature review). At the same time, open sourcing frontier models accelerates experimentation in ways that proprietary vendors cannot easily control.

Regulatory tensions and geopolitical friction

All of these technical and commercial shifts intersect with policy. Several regulators and governments have already flagged DeepSeek’s data handling and national‑security profile. European authorities have investigated and in some cases ordered temporary blocks or app removals, and a range of governments have advised caution or restricted use on official devices. Those actions complicate adoption in regulated sectors and underline that open availability of weights does not remove concerns about data flows or access by foreign governments.

Companies contemplating deploying these models need to think about data residency, compliance with local privacy rules, and supply‑chain provenance for training and inference hardware—issues that are now central to procurement and risk assessments rather than technical afterthoughts.

What this means for the AI landscape

There are three broad takeaways. First, architectural efficiency (not just brute‑force scale) can move the frontier, especially for long‑context and agentic tasks. Second, open release of high‑capability models forces incumbents to rethink pricing and product strategy: governments, enterprises and developers now have an alternative that is easier to self‑host. Third, policy and trust remain gating factors—technical progress alone won't determine who wins or how widely these systems are deployed.

For European and U.S. organisations in particular, the challenge is practical: balance the operational and cost benefits of a freely available, efficient model against unresolved questions about data governance, third‑party audits, and regulatory risk. The next several months will be a live experiment in how the market, regulators, and providers adapt.

What I’ll be watching

• Independent audits and replication of DeepSeek’s benchmark claims.
• Enterprise term‑sheets showing who chooses to self‑host these weights and under what safeguards.
• Regulatory rulings that clarify how data‑protection rules apply to foreign‑hosted model services and open weights.
• How major cloud and silicon vendors respond—both technically (runtime support, optimised kernels) and commercially (pricing, partnerships).

DeepSeek’s release is a reminder that the AI race is now about multiple levers—architecture, data, tooling, distribution and regulation—not just raw compute. For engineers, product leaders and policymakers, that complexity is a feature: it creates both opportunity and a lot of hard questions to answer before these capabilities become foundational infrastructure.

— Mattias Risberg, Dark Matter