DeepSeek-R1- a local thinking model

Regardless of which benchmark you look at, the DeepSeek-R1 is an impressive family of models that have been completely open sourced. They just rolled out DeepSeek-V3, a model that can code as well as Sonnet 3.5 about 3 weeks ago.

Here’s what Jim Fan, one of the top researches in AI has to say about DeepSeek-R1.

If you run your own local models using Ollama and have an NVIDIA RTX-4090, you can easily load up the 32b parameter version which has been highlighted on this chart.

ollama run deepseek-r1:32b

You can see the raw model thinking about the problem and considering and ‘searching’ down the path of various outcomes that seem reasonable. In the raw output you can see the think tags, these are the tokens being spent to ground and search. This is what ‘test-time scaling’ and ‘chain of thought’ buzzwords are all about. It is eerily reminiscent of how we think to ourselves as you read through. There’s a good reas

The full DeepSeek-R1, the whole package is all MIT licensed and thus commercial friendly. The main model may not be able to be run in some environments, but they have provided distillation examples using llama and qwen models (US and China based based models). Several of these distilled models rival and get close to the main model’s capabilities.

The important part of that release is that it paves the way for others to distill their own thinking models to think in new ways and directions using the same process - taking off the shelf llama models, applying the deepseek thinking training mechanisms (all outlined in this research), and training your own ‘o1’ level model for a fraction of the cost that OpenAI spent on it.

Open Source AI is about 3 months behind the frontier models it seems.

Bottom Line:

**Performance **: DeepSeek-R1 is stated to have performance comparable to OpenAI-o1

**Open-source **: The model and technical report are fully open-source under MIT license, allowing for free distillation and commercialization.

**Transparent thinking **: It includes a transparent thought process in real-time, with upcoming open-source models and API availability.

**Driving Costs Down : **

Conclusion

There’s a good reason the industry is abuzz about this release, it marks a moment where in the last month both OpenAI’s flagship o1 and Anthropic’s Sonnet 3.5 have been challenged by DeepSeek - R1 and V3 respectively.

DeepSeek V3 is not available on consumer hardware, and thus that was big news, but a lot of folks did not pay attention to it. However, with this release, DeepSeek R1, most consumers can run a slightly quantized version of it locally on an RTX-4090 at decent speeds.