Gemini 3 Flash, at $0.50 per million input tokens, hits a sweet spot for performance and low cost. I don’t know yet, but I might not have to use the Pro API, which is 4 times as expensive, very often.

The new 30B model nvidia/nemotron-3-nano is very good with tool use - I just checked it using LM Studio against all of my tool and localMCP server projects. A pleasant surprise: I have a 32G M2-Pro Mac system, and the model weights are over 20G, but performance is quite good -

I think Gemini will overtake ChatGPT. My non-tech brother spent a few minutes on the phone yesterday raving about how he uses Gemini on his new Pixel phone for just about everything. I am curious to see how good the Gemini iOS integration will be next year. One more advantage: I

After Wave was dropped, I still played around with the open source release for a year - a major waste of time and I had problems understanding and slightly modifying the code base, but I had a lot of fun with it, so not a flop for me! re: grandparent comment about AntiGravity

I have a project that I want to work on next year - my goal is not to build anything practical, rather to experiment with ideas. I am trying to decide whether to use Python, Common Lisp, or Clojure. I stood up a project web site as a placeholder:

At the command line, I use (ql:quickload …) 95% of the time and (load …) 5% of the time. In my .sbclrc file I alias ‘ql’ to ‘ql:quickload’. In Emacs I open a REPL and manually (ql ….) a project, then just evaluate expressions as I

OK, running Mistral vide with local ‘ollama run devstral-small-2’ is really too slow (but works OK). Running ollama on command line with a detailed prompt and copying code to be modified is more practical:

Hey! devstral-small-2 is a very good coding model. I don’t have it hooked up to Mistral Vibe yet, just running with ‘ollama run devstral-small-2’ copying in some of my older Clojure and Python code and asking for improvements. All modified code ran tests correctly. Amazing for a

Well, I feel better informed now than I did an hour ago before I read @Tim_Dettmers article:

For a dense 8B model rnj-1 runs quickly on my Mac. I experiment with my own low level (no dependencies) agentic code so fast local execution is important:

So nice: rnj-1 from essentialAI. I have tested it on a variety of tool use cases and asked for Python code generation for a graph theory problem. 100% pass on my tool use test cases and the 100 lines of generated Python code worked perfectly. Run with: 'ollama run rnj-1’ or use

Some more information: One working Mistrail-3:3b tool example uses the standard Ollama Python setup (tools=[local_files] inside https://t.co/ckgdiUgB8W). My second example uses tools from scratch using only the Python requests library:

I used the Ollama Python library for:

I enjoy writing code from scratch, especially when experimenting with tool use and LLM-based agents. I am pleasantly surprised at how well ministral-3:3b does for tool use - probably a priority training the model.

Bravo @MistralAI ! Thank you for the Apache 2.0 license! And the MoE and the multimodal! Everything I've been wishing for is delivered. A thinking multimodal vision model! I will start mapping the refusal space. (Compared with Qwen3)

Mistral is back to their roots All apache 2.0 Strong models, multimodal, multiple types, and MORE BASE MODELS Huge W, smart by them to capitalize on the power vacuum of companies looking for alternates to Chinese models. So happy Chinese companies are pushing change globally.

For a 8B parameter model mistral-3:8b seems very good, testing it now with my collection of 'oddball' private test prompts. I have seen some negative comments on the new Mistral releases but for small embedded and local applications Mistral is in the same ballpark as other 2025

Gary, I enjoyed your last book and just signed up for your newsletter. I for one am happy to see more basic non-LLM research (even though I spend almost an hour a day using LLM based systems as tools).