Tech Grandpa

Ok, this is exciting! DeepResearch, but OpenSource. tongyi-agent.github.io/blog/introdu...

Moonshot AI has released the updated Kimi K2-0905

- Enhanced coding capabilities, esp. front-end & tool-calling
- Context length extended to 256k tokens
- Improved integration with various agent scaffolds

DeepSeek-V3.1 - a deepseek-ai Collection We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Does anyone know if this is the new DeepSeek model? No model card, no benchmarks so far, but unsloth already made a quant 😊 huggingface.co/collections/...

More #gpt-5 + #cursor impressions...
I tried to fix a bug while running 80-90% context size and it basically circled around the same ideas, no matter what I told it for an hour back and forth. As the context got too big, I started a new chat with the "auto" mode (by accident) and it one-shotted it.

#GPT-5 Take away:
- Coding is ok on medium size code bases
- Tool calling seems great, but I haven't tested it enough to tell
- I miss the variety of models before
- the price is a huge win for #openai - people don't yet understand the impact
- the router is annoying, but something others will adapt

Well, the gpt-5 launch was ... interesting. I spent the last days testing it out in cursor and I am not impressed unfortunately. The thinking was taking many loops and working on profanities. The UI design were also not winning prices - not bad, don't get me wrong, just not great.

🤣🤣🤣

Source: x.com/avichal/stat...

Great comparison of local LLMs and their performance on consumer grade cards (24GB RAM limit):
www.reddit.com/r/LocalLLaMA... #AI #LLM #homelabai #localaiagent

Hmm, didn’t know Kimi, but would have it expected to solve the strawberry challenge given its popularity. Probably, there are other properties than performance that convince people. Would be interesting to know what makes people choose one over the other - relatability of conversational style?

deepseek-r1:14b DeepSeek's first generation reasoning models with comparable performance to OpenAI-o1.

I tested the new DeepSeek.r1 14b, which performs pretty well and is a good middle ground between speed, VRAM consumption and quality. It fails the “strawberry” test, but I can live with that 😉. If you can afford it, go with the 70b model though.

ollama.com/library/deep...

A short guide to run DeepSeek R1 (all 671B of it) on a home cluster of Macs with mlx.distributed.

gist.github.com/awni/ec071fd...

YouTube video by The Nitty-Gritty Will Project DIGITS Revolutionize The Way We Use LLMs Forever?

I just posted a video about #NVIDIA and its project DIGITS. I not only think that it is a marvel of engineering to put so much #ai power into such a tiny package, but I also think that this is the beginning of a revolution: Edge AI computing for the masses. Check it out:
youtu.be/NEi9oJbwZC4

GitHub - deepseek-ai/DeepSeek-R1 Contribute to deepseek-ai/DeepSeek-R1 development by creating an account on GitHub.

In case you have been living under a rock, deepseek has released its new r1 thinking model rivaling OpenAI’s o1 family, but being open source and MIT licensed! Ollama and HF already provide quants and distilled versions! Great times! github.com/deepseek-ai/...