Please join our Discord server! https://discord.gg/XCazaEVNzT

Simon Willison s Weblog

From Speedrunwiki.com
Revision as of 08:55, 23 February 2025 by Milla12M313 (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigationJump to search


That model was trained in part utilizing their unreleased R1 "thinking" design. Today they have actually R1 itself, in addition to an entire family of new models obtained from that base.


There's a great deal of things in the new release.


DeepSeek-R1-Zero seems the base model. It's over 650GB in size and, like the majority of their other releases, is under a clean MIT license. DeepSeek warn that "DeepSeek-R1-Zero comes across challenges such as unlimited repetition, bad readability, and language blending." ... so they likewise launched:


DeepSeek-R1-which "incorporates cold-start information before RL" and "attains performance comparable to OpenAI-o1 across math, code, and thinking tasks". That one is likewise MIT licensed, and is a comparable size.


I don't have the capability to run designs bigger than about 50GB (I have an M2 with 64GB of RAM), so neither of these two models are something I can easily have fun with myself. That's where the brand-new distilled models are available in.


To support the research study neighborhood, we have open-sourced DeepSeek-R1-Zero, DeepSeek-R1, and 6 dense designs distilled from DeepSeek-R1 based on Llama and Qwen.


This is a remarkable flex! They have models based upon Qwen 2.5 (14B, 32B, Math 1.5 B and Math 7B) and Llama 3 (Llama-3.1 8B and Llama 3.3 70B Instruct).


Weirdly those Llama models have an MIT license connected, which I'm uncertain is compatible with the underlying Llama license. Qwen models are Apache certified so perhaps MIT is OK?


(I also simply saw the MIT license files say "Copyright (c) 2023 DeepSeek" so they may need to pay a little bit more attention to how they copied those in.)


Licensing aside, these distilled designs are fascinating beasts.


Running DeepSeek-R1-Distill-Llama-8B-GGUF


Quantized variations are already starting to appear. Up until now I've attempted simply among those- unsloth/DeepSeek-R 1-Distill-Llama-8B-GGUF launched by Unsloth AI-and it's really enjoyable to have fun with.


I'm running it utilizing the mix of Ollama, LLM and the llm-ollama plugin.


First I brought and ran the design using Ollama itself:


This downloads around 8.5 GB of design data and starts an interactive chat interface.


Once the design has actually been brought, LLM can talk to it as well. I choose using LLM for experiments because it logs everything to SQLite for later exploration.


I installed the plugin and ran it like this:


Each response begins with a ... pseudo-XML tag containing the chain of thought utilized to help create the action. For "a joke about a pelican and a walrus who run a tea room together" that started like this:


Okay, so I need to come up with a joke about a pelican and a walrus running a tea room together. Hmm, that's an intriguing mix. Let me think of how these 2 characters may communicate in an amusing situation.


First, let's consider their attributes. Pelicans are known for their long beaks and Webbed feet, frequently seen near the beach or water. Walruses have big teeth, thick fur, and they're generally discovered in cooler environments, like icebergs or snowy areas. So, integrating these two into a tea space setting is already a funny image.


[...]

It then believed for 20 paragraphs before outputting the joke! Here's the complete transcript and here's the output at the end, after the tag:


The Joke:


A pelican and a walrus choose 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 mishap, overturning the teapot. Meanwhile, the pelican, trying to assist, uses its beak to position dishes on the table, triggering a couple of spills.


After a series of comical mishaps, the walrus searches for and states with a grin, "This isn't so fishy anymore." The pelican smirks and says, "Maybe not, however we do have a lot of krill in our tea!"


Explanation:


- Setup: Pelican and Walrus run an unlikely tea space. Their sizes lead to comedic chaos.
- Punchline: "This isn't so fishy any longer" is a pun on the scenario ending up being less ridiculous. The follow-up joke presents "krill," a play on words that ties into both marine life ("fish") and tea, making it funny while fitting the style.


So yeah, the joke is horrible. But the process of arriving was such an interesting insight into how these brand-new models work.


This was a fairly small 8B design. I'm looking forward to trying the Llama 70B variation, which isn't yet available in a GGUF I can run with Ollama. Given the strength of Llama 3.3 70B-currently my preferred GPT-4 class design that I've worked on my own machine-I have high expectations.


Update 21st January 2025: I got this quantized variation of that Llama 3.3 70B R1 distilled model working like this-a 34GB download:


Can it draw a pelican?


I attempted my classic Generate an SVG of a pelican riding a bike timely too. It did refrain from doing extremely well:


It aimed to me like it got the order of the components incorrect, so I followed up with:


the background ended up covering the remainder of the image


It believed some more and gave me this:


Similar to the earlier joke, the chain of thought in the records was far more intriguing than completion outcome.


Other methods to attempt DeepSeek-R1


If you want to attempt the model out without setting up anything you can do so using chat.deepseek.com-you'll require to produce an account (sign in with Google, use an email address or offer a Chinese +86 contact number) and after that pick the "DeepThink" option below the timely input box.


DeepSeek use the model via their API, using an OpenAI-imitating endpoint. You can access that via LLM by dropping this into your extra-openai-models. yaml setup file:


Then run llm keys set deepseek and speedrunwiki.com paste in your API key, then utilize llm -m deepseek-reasoner 'timely' to run triggers.


This won't reveal you the thinking tokens, unfortunately. Those are served up by the API (example here) but LLM does not yet have a way to show them.