Gemma 4 is here. 🚀
Google DeepMind’s new open model family — Apache 2.0, multimodal, and capable enough to run serious workloads on your own hardware.
We at 575 Lab couldn’t wait to test it.
Our full review below
Curious to hear what others are seeing in their own tests.
#Gemma4 #AI #OpenSource
MiniMax has release the MiniMax M2.7 model weights.
Blog: www.minimax.io/news/minimax...
Model: huggingface.co/MiniMaxAI/Mi...
Really appreciate the thoughts from Nathan. I prefer open over closed source models, but Nathan is raising interesting points.
Great read. Thank you for sharing @natolambert.bsky.social
People are too obsessed with benchmarks for open models. The core determining factor of success often is:
1. Immediate & long term tooling support
2. Finetunability
Tbh Gemma has struggled here in the past. Qwen has excelled at it. It's where the winners are crowned.
So excited about Gemma4!
Nemotron-3-Ultra and Gemma4!
Mistral's Voxtral TTS
An open-weight model for natural, expressive, and ultra-fast text-to-speech
🎭Realistic, emotionally expressive speech.
🌍Supports 9 languages and accurately captures diverse dialects.
⚡Very low latency for time-to-first-audio.
🔄Easily adaptable to new voices
Kolmogorov-Arnold causal generative models
KaCGM, a causal generative model for mixed-type tabular data where each structural equation is parameterized by a Kolmogorov--Arnold Network (KAN). This decomposition enables direct inspection of learned causal mechanisms,
I'm a big fan of Nemotron models
Mostly because the API is free 😂
PaperBanana
AI researchers often spend way too much time crafting diagrams and plots instead of focusing on the ideas. To rescue them from this burden, they built an Agentic Framework to auto-generate NeurIPS-quality paper illustrations!
Dash
Inspired by OpenAI’s in-house data agent ( openai.com/index/inside... ), it is a self-learning data agent that grounds its answers in 6 layers of context and improves with every run.
github.com/agno-agi/dash
🧠🆕 NEW: Deploying Small Language Models (LFWS307) — a live 3-day workshop with 10+ hrs of hands-on labs. Learn to deploy SLMs on laptops, servers, edge & browser. Build skills for real-world AI & ML platform roles.
Enroll: training.linuxfoundation.org/training/dep...
#SLMs #AIEngineering #MLOps
Agentic AI is more than just a tool. LF executive director Jim Zemlin explains why smaller, specialized AI agents deliver better results, lower costs, and greater trust with open source.
🎥https://www.youtube.com/watch?v=vfNMZ2G8jPA&t=254s
#OpenSource #AI #AgenticAI #Innovation #ResponsibleAI
BREAKING:
TIM WALZ: "If you see these ICE agents in your neighborhood, take out that phone and hit record."
"Help us create a database of the atrocities against Minnesotans, not just to establish a record for posterity, but to bank evidence for future prosecution."
A sentiment I very much share. I hope not to get disappointed by @anthropic.com
Llama a la Coalición de Derechos de los Inmigrantes de Portland (1-888-622-1510) si ves a ICE en su vecindario, y llama a mi oficina (503-231-2300) si un ser querido ha sido detenido.
Ahora más que nunca, debemos proteger a nuestros vecinos. Alza la voz.
2026 is the Year of Self-hosting by Jordan Fulghum
CLI agents like Claude Code make self-hosting on a cheapo home server dramatically easier and actually fun.
fulghum.io/self-hosting
Please welcome Google's Open Source efforts to Blue Sky at @opensource.google!
someone saying, "You cannot afford to run an open source model on your own hardware" in response to me saying I advocate for open source AI
The post they are replying to was written on an MacBook Pro I upgraded to 64 GB of RAM specifically to be able to run 70B models
You can run Qwen3-Coder-30B on 32GB or RAM, and it's surprisingly capable at most coding tasks, even
@maxnum.bsky.social macht Mut. Hoffnung, keine Angst!
[YouTube] Mastering the JAX AI Stack
The Learning JAX video series is now live! 23 videos covering the entire JAX AI Stack to an intermediate/advanced level www.youtube.com/playlist?lis...
Humans really had the creative freedom to do anything and we chose racism and credit scores
![mkdir /tmp/aime25_openai
OPENAI_API_KEY=x \
uv run --python 3.13 --with 'gpt-oss[eval]' \
python -m gpt_oss.evals \
--base-url http://localhost:1234/v1 \
--eval aime25 \
--sampler chat_completions \
--model openai/gpt-oss-20b \
--reasoning-effort low \
--n-threads 2](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:kft6lu4trxowqmter2b6vg6z/bafkreib5inem3k2p7bv3sepg4w56r5a7vzsfcj6ilmvvvr6bgzrz6lu4wq)
mkdir /tmp/aime25_openai OPENAI_API_KEY=x \ uv run --python 3.13 --with 'gpt-oss[eval]' \ python -m gpt_oss.evals \ --base-url http://localhost:1234/v1 \ --eval aime25 \ --sampler chat_completions \ --model openai/gpt-oss-20b \ --reasoning-effort low \ --n-threads 2
TIL you can run an eval suite against OpenAI's gpt-oss-20b open weights model running in LM Studio with the following uv one-liner. Here's my full TIL: til.simonwillison.net/llms/gpt-oss...
Print featuring the upper half of a white woman in a dress with hair up facing forwards
Fanny Mendelssohn (1805-1847)
composed more than 460 works, which were published under her lauded composer brother, Felix Mendelssohn's name #WomensArt
I remember trying to deploy Llama1. Amazing how the ecosystem has changed!
Der öffentlich-rechtliche Rundfunk als Plattform für eine diverse und lebendige Zivilgesellschaft. Mein Plädoyer auf dem Monitor-Forum der re:publica. Mit Annika Brockschmidt, Constanze Kurz und Nadia Zaboura.
youtube.com/watch?v=C2_L...
Ausführlicher Rückblick auf die @re-publica.com bei ‹Breitband› im Deutschlandfunk Kultur: «Digitaler Faschismus und was dagegen hilft»
#rp25
www.deutschlandfunkkultur.de/re-publica-2...
BIG news for Gemma this week. Gemma 3n, MedGemma, and more. Watch Omar Sanseviero and I to learn what's new
Watch the full session →
youtu.be/o7Bv4r08FBM?...
