I’ve been spending some time with Deepseek R1 over the last few days after it became available on Ollama.
In summary, it’s a capable reasoning model but it is a product of its geography.
The screenshots below show clean runs of the model, being asked about particularly politicised events in China and US history. Deepseek R1, even under repeated prompting won’t talk about Tiananmen Square at all, but it’s quite happy to talk through the January 6 insurrection without any censorship.
Even with some light steering, the model refuses to answer
No qualms working through this historical event…
A reasonable summary of events given no lookups allowed.
This isn’t really surprising given the provenance of Deepseek, and for many business use cases, such as a summarisation or decisioning engine, this won’t matter at all. But, because of its inherent safety controls relating to banned topics, the edges of this are likely to bleed over into a number of adjacent topics and cause weird behaviour.
Imagine being denied a summary when asking about Winnie the Pooh for instance. In some cases this would be inexplicable unless you have the context for why this might be banned.
This becomes something else you need to be aware of when using these models (on top of all of the other “normal” adherence, hallucination and error issues). I’d be wary about using them to create synthetic data because of this too.
For me, the key story about Deepseek R1 and Deepseek V3 are that open models are closing the gap on the big closed ones and this builds on the trajectory we’ve already seen from Llama3.2, Qwen Coder, Phi4, and others.
There are no moats when it comes to foundational models - the only moat will be bundling these capabilities up into a compelling UX / CX to make them easy to use in orgs that want to deploy them. This is the same strategy that allowed the big Cloud Tech companies to take off (for both good and bad impacts on the Open Source community - but that’s another post).
We’re still very much in the “A computer is the size of a fridge” period of AI models so smaller, faster, cheaper to train and run is very much the future and this is another data point along that path.