That model was trained in part utilizing their unreleased R1 "thinking" model. Today they've launched R1 itself, along with a whole household of brand-new designs obtained from that base.
There's a lot of stuff in the new release.
DeepSeek-R1-Zero seems the base design. It's over 650GB in size and, like most of their other releases, is under a tidy MIT license. DeepSeek caution that "DeepSeek-R1-Zero comes across obstacles such as limitless repetition, poor readability, and language mixing." ... so they likewise released:
DeepSeek-R1-which "includes cold-start information before RL" and "attains efficiency comparable to OpenAI-o1 throughout mathematics, code, and thinking tasks". That a person is also MIT accredited, and is a comparable size.
I don't have the capability to run models bigger than about 50GB (I have an M2 with 64GB of RAM), so neither of these two models are something I can quickly have fun with myself. That's where the brand-new distilled models are available in.
To support the research study community, we have open-sourced DeepSeek-R1-Zero, DeepSeek-R1, and six thick models distilled from DeepSeek-R1 based upon Llama and Qwen.
This is a remarkable flex! They have designs based on Qwen 2.5 (14B, utahsyardsale.com 32B, Math 1.5 B and Math 7B) and securityholes.science Llama 3 (Llama-3.1 8B and Llama 3.3 70B Instruct).
Weirdly those Llama models have an MIT license attached, which I'm uncertain works with the underlying Llama license. Qwen designs are Apache licensed so perhaps MIT is OK?
(I likewise simply noticed the MIT license files state "Copyright (c) 2023 DeepSeek" so they may require to pay a bit more attention to how they copied those in.)
Licensing aside, these distilled designs are interesting beasts.
Running DeepSeek-R1-Distill-Llama-8B-GGUF
Quantized variations are already starting to appear. Up until now I've tried simply among those- unsloth/DeepSeek-R 1-Distill-Llama-8B-GGUF launched by Unsloth AI-and it's really fun to play with.
I'm running it using the mix of Ollama, LLM and the llm-ollama plugin.
First I brought and ran the model using Ollama itself:
This downloads around 8.5 GB of design information and starts an interactive chat interface.
Once the model has been brought, LLM can talk to it too. I choose utilizing LLM for experiments because it logs everything to SQLite for later exploration.
I set up the plugin and ran it like this:
Each response starts with a ... pseudo-XML tag containing the chain of idea used to assist create the reaction. For "a joke about a pelican and a walrus who run a tea room together" that started like this:
Okay, so I require to come up with a joke about a pelican and a walrus running a tea room together. Hmm, wiki.myamens.com that's an interesting mix. Let me think of how these 2 characters may communicate in a funny situation.
First, let's consider their qualities. Pelicans are known for their long beaks and Webbed feet, often seen near the beach or water. Walruses have huge teeth, thick fur, ratemywifey.com and they're normally found in chillier climates, like icebergs or snowy locations. So, these 2 into a tea room setting is currently a funny image.
[...]
It then thought for 20 paragraphs before outputting the joke! Here's the full transcript and here's the output at the end, after the tag:
The Joke:
A pelican and a walrus decide to open a quaint little tea space together. The walrus, with its huge size, struggles to find comfortable chairs, so it sits on the table by accident, knocking over the teapot. Meanwhile, the pelican, trying to assist, uses its beak to position saucers on the table, triggering a few spills.
After a series of comical accidents, the walrus searches for and says with a grin, "This isn't so fishy any longer." The pelican smirks and says, "Maybe not, however we do have a great deal of krill in our tea!"
Explanation:
- Setup: Pelican and it-viking.ch Walrus run a not likely tea room. Their sizes lead to comedic turmoil.
- Punchline: "This isn't so fishy any longer" is a pun on the circumstance becoming less unreasonable. The follow-up joke presents "krill," a play on words that connects into both marine life ("fish") and tea, making it amusing while fitting the style.
So yeah, the joke is awful. But the process of arriving was such a fascinating insight into how these new designs work.
This was a fairly small 8B design. I'm eagerly anticipating experimenting with the Llama 70B variation, which isn't yet available in a GGUF I can keep up Ollama. Given the strength of Llama 3.3 70B-currently my preferred GPT-4 class design that I've run on my own machine-I have high expectations.
Update 21st January 2025: I got this quantized version of that Llama 3.3 70B R1 distilled design working like this-a 34GB download:
Can it draw a pelican?
I attempted my classic Generate an SVG of a pelican riding a bicycle prompt too. It did refrain from doing extremely well:
It aimed to me like it got the order of the components incorrect, kenpoguy.com so I followed up with:
the background wound up covering the remainder of the image
It believed some more and offered me this:
As with the earlier joke, the chain of thought in the records was much more interesting than the end result.
Other methods to attempt DeepSeek-R1
If you desire to try the design out without installing anything you can do so utilizing chat.deepseek.com-you'll require to produce an account (check in with Google, use an email address or offer a Chinese +86 contact number) and after that pick the "DeepThink" alternative listed below the prompt input box.
DeepSeek use the model via their API, utilizing an OpenAI-imitating endpoint. You can access that by means of LLM by dropping this into your extra-openai-models. yaml setup file:
Then run llm keys set deepseek and paste in your API secret, then utilize llm -m deepseek-reasoner 'prompt' to run prompts.
This will not reveal you the thinking tokens, sadly. Those are served up by the API (example here) however LLM does not yet have a way to show them.