Open-Source AI Just Caught Up

Early 2025. A Chinese lab most people had never heard of releases a reasoning model that matches OpenAI's o1 in math and coding. Reported cost: roughly $5.3 million in compute for the final training run. That figure covers GPU hours only, not R&D, not failed experiments, not the $51M+ hardware investment. Frontier models from the big labs reportedly cost tens of millions. Sometimes more.

DeepSeek R1 did it with 2,048 GPUs where others use 16,000-plus. And they open-sourced the whole thing under MIT license.

The closed-model dominance everyone assumed was permanent just took a hit.

What DeepSeek R1 actually did

The benchmarks matter here, so let's be precise.

DeepSeek R1 hit 79.8% on AIME 2024 (the American Invitational Mathematics Examination), matching OpenAI o1's 79.2%. It also hit 97.3% on MATH-500. These are the reference benchmarks the industry uses to measure reasoning capability, not cherry-picked obscure tests.

The model learns to reason through reinforcement learning, no human-written reasoning examples required. DeepSeek documented the full approach. Anyone can study it or replicate it.

They also released six smaller distilled versions based on the Llama and Qwen architectures. The 32B version (DeepSeek-R1-Distill-Qwen-32B) outperforms OpenAI's o1-mini on multiple evaluations. A 32-billion-parameter model you can run on your own hardware, beating a commercial product.

It's not just DeepSeek

"Open source is catching up" claims have come and gone before. What's different this time is that it's happening on multiple fronts at once.

Qwen from Alibaba keeps accelerating. The Qwen 2.5 family was already competing with GPT-4o. Since then Alibaba shipped Qwen 3 (April 2025) with hybrid reasoning across dense and MoE architectures, then Qwen 3.5 (February 2026) at 397 billion parameters with 201-language support. All under permissive licenses.

Mistral Small 3.1 landed in March 2025. 24 billion parameters, runs on a single RTX 4090 or a Mac with 32GB RAM. 150 tokens per second, 128K context, image understanding. Apache 2.0 license. Two years ago this level of capability needed a cluster. Now it runs on a laptop.

DeepSeek V3 keeps evolving. V3-0324 folded in reasoning improvements from R1. V3.2 (December 2025) pushed further still. The iteration speed is remarkable.

Why this matters beyond benchmarks

I could keep stacking numbers. The real story is what these models make possible.

Self-hosting becomes viable

When the best open models sat clearly below GPT-4, the self-hosting case was hard to argue. Private and cheaper, sure, but the quality gap weighed too much. That argument is collapsing.

DeepSeek R1 handles reasoning. Qwen 3 covers general capability. Mistral Small delivers efficient inference. These models are good enough that the quality tradeoff turns marginal for a lot of real-world use cases. In return: full data control, no per-token billing at scale, no dependency on a provider's uptime or policy changes.

For regulated industries, healthcare, finance, legal, running a frontier-quality model entirely inside your own infrastructure changes the equation. You can't send patient records to an API endpoint. You can run inference on your own servers.

The cost equation flipped

DeepSeek reports $5.3 million in compute for R1's base model. Compare that against what a commercial reasoning API costs at scale. For a lot of companies, self-hosting an open model pays for itself within months.

This isn't theoretical. I've helped clients run the break-even numbers. Past a few thousand API calls a day, the math starts favoring self-hosted open models. Add the latency gains from local inference, no network round-trip, and the case gets stronger still.

The talent pool widens

Once a model is open, researchers and engineers everywhere can pick it apart, fine-tune it for specific tasks. DeepSeek's RL training technique is already being adapted by other labs. The distilled versions run on consumer hardware. A solo developer can experiment with reasoning-capable AI.

This is how Linux won the server market. Not one company, thousands of contributors stacking improvements over years.

What I check in an open model

Not all open-source AI is equal. Here's what I look at.

License first. MIT and Apache 2.0 mean you can use the model commercially without restriction. Some "open" models carry non-commercial clauses or custom licenses that limit real use. DeepSeek R1 is MIT. Mistral Small 3.1 and Qwen2.5-VL are Apache 2.0. Those are genuinely open.

Then reproducibility. DeepSeek shared its training methodology. That's rare and it's worth a lot: the community can verify results, spot weaknesses, build on top. A model released without training details is harder to trust.

Hardware requirements matter too. A 671-billion-parameter model might post great results, but if it needs eight A100s to run, it's "open" only for companies with serious GPU budgets. The distilled and small models, DeepSeek-R1-Distill at 32B, Mistral Small at 24B, that's where the practical value sits.

And fine-tuning. Can you adapt the model to your domain? Open weights are table stakes. Open training code and clear documentation are what separate a downloadable demo from a production tool.

The practical path forward

Running a business and wondering what to do with all this?

Don't wait for "the best" model. There won't be one. The lineup changes every few weeks. Pick a model good enough for your use case, deploy it, improve from there. Qwen 3 and Mistral Small are solid general picks.

Start with a specific use case. "We want to use open-source AI" isn't a strategy. "We want to automate internal document summaries without sending data outside" is one.

Plan your infrastructure. Running models locally needs GPU compute. If you don't have it in-house, dedicated inference providers host open models for a fraction of API prices. The middle ground between "fully self-hosted" and "send everything to OpenAI" is wider than most people think.

Keep watching. Meta's Llama 4 shipped in April 2025 with the multimodal Scout (109B) and Maverick (400B) models. Qwen 3 brought hybrid reasoning to open weights. The pace isn't slowing down. What's frontier today is baseline in six months.

If you're looking to bring open-source LLMs into your workflows, now's the time. The models are capable, the tooling is maturing, the cost advantage is real. Open-source AI has crossed the line from interesting experiment to credible production option. The only question left is whether you build on open foundations or stay locked into closed APIs.

What's next?

The tortoise catches the hare

For the fundamentals on AI agents: start here. And to host these models yourself, check the practical guide.