Stellan Haglund

@ivanfioravanti What did you reach with 0.6B?

Playing with my italian wine classification and Qwen3.5-0.8B fine-tuning with MLX on Apple Silicon! From 10% to 72% after first 100 iters with batch size 12. Let's see what I can reach with it!

@mkurman88 Seeing a lot of people saying 4b is great to

@Stellanhaglund I jumped straight into 9B, and it’s a goat!

Qwen 3.5 4B !!!! 🌶️🌶️🌶️

@elonmusk @tetsuoai Agree, I take it it’s ok to do this from @grok then?

@tetsuoai Banger 🤣🤣 How dare they steal the stuff Anthropic stole from human coders??

I can't believe someone would just steal from Anthropic like this. The millions of man-hours Anthropic spent hand-writing code, text, art, books, etc. to generate enough data for training must be taken into consideration here. Where is the respect for IP?

@OfficialLoganK can I use a fine tuned version of flash 2.0 with the batch api?

@OfficialLoganK Does it work with fine tuned models of either version?

We just shipped "Batch mode" in the Gemini API with 50% discounts on our 2.5 models and the ability to enqueue billions of tokens at a time!

@_philschmid @multimodalart A bit out of context, but I’ve been trying to find an answer to if it’s possible to use a fine tuned model with the batch prediction api?

@multimodalart Sorry it is: - Gemini 2.5 Flash (Stable, updated pricing from 05-20)

Gemini 2.5 is production ready! We just launched 3 new Gemini models with 2.5 Pro and Flash being now generally available and a new Gemini 2.5 Flash Lite preview! 🧠⚡️🔦 Here is all you need to know: 🔦 New Gemini 2.5 Flash Lite (Preview) with Thinking, 1M context, only $0.1/$0.4, better as 2.0 flash and tool-use. 🧠 ⚡ Gemini 2.5 Flash and 2.5 Pro are now Generally Available (GA) and production-ready. 💰 Updated pricing for 2.5 Flash to $0.30/1M input and $2.50/1M output tokens. 🤗 Start building all 3 Models today in AI Studio and via Gemini API! New Model IDs: `gemini-2.5-flash-lite-preview-06-17`, `gemini-2.5-flash`, `gemini-2.5-pro`

@LingYang_PU That looks pretty sequential, isn’t the upside with diffusion to predict all the tokens every step?

We present MMaDA, first diffusion that unifies text reasoning, multimodal understanding, and image generation through Mixed Long-CoT, and unified RL - UniGRPO. 📚 Paper: arxiv.org/abs/2505.15809 💻 Code: github.com/Gen-Verse/MMaDA 📦 Model: huggingface.co/Gen-Verse/MMaD…

@victormustar How much memory does it need?

You can really feel the A3B here 🚀 (hardware is mbp m4 max 128GB btw)

Qwen/Qwen3-30B-A3B is absolute game changer 🤯 ⬇️ Are you on a Mac? then use MLX weights - getting 100 tokens/sec here. It changes everything for real world use cases.

@rodydavis Do you also compile all these to wasm?

I was able to create a CRDT in SQLite with pure C (no dependencies) custom loadable extensions! 🚨 uuid.c (From official SQLite misc extensions) hlc.c (Dart HLC ported to C) crdt.c (Dart logic with triggers ported to C) gist.github.com/rodydavis/d197…

@steren @ollama Tokens per second?

With @ollama 0.6.2, Gemma 3 27B now runs on 1 Cloud Run GPU. That's 27 billion parameters. The biggest Gemma 3 variant. The most capable open model you can run on a single GPU.

@OfficialLoganK is it possible to use masks with flash 2 image generation somehow?

@OfficialLoganK Can you add masks?

We just updated the Gemini API docs to make native image generation 🖼️ even more simple to access and get started with across Python, JS, and cURL, happy building : ) ai.google.dev/gemini-api/doc…

@piersmorgan So what’s your solution?

This is good to see, but nobody, including President Zelensky, should be under any illusion that Ukraine needs America to end the war and get peace.

@piersmorgan It’s a complete disaster

Blocking so many loons tonight. Is it just me or is X getting madder by the day?

@abacaj Really like your recent content on grpo! 👍

GRPO llama3.2-1B inject "Wait a second..." when model is wrong and let it regenerate then compute loss

@cerebras Like is it possible to make sure the data doesn’t leave Europe when using the api?

@Stellanhaglund Flash Answers can be disabled in settings

@cerebras what are the options for European customers with GDPR requirements?

@cerebras I mean for inference, the data centers are located in the us right?

@elonmusk when something is community noted, why aren’t other posts containing the same argument also marked with the same note? I see posts all day saying stuff I already saw was community noted as false several days ago.