Olmo-core 3: AI2's Open Bet on Trillion-Parameter MoE Training
AI2 just released the training stack behind next-gen Olmo. It scales mixture-of-experts to 1.2T parameters while keeping throughput high—and it's fully open.
#…
Loading…
15 posts
AI2 just released the training stack behind next-gen Olmo. It scales mixture-of-experts to 1.2T parameters while keeping throughput high—and it's fully open.
IBM's new Granite Time Series PatchTST-FM-r2 hits #1 among commercially licensed zero-shot forecasters on GIFT-Eval. 385M params, conformer blocks, and actual permissive licensing.
A developer frustrated with Copilot bills built a deterministic terminal assistant in Python that translates natural language to shell commands—no embeddings, no ML, millisecond responses.
IBM just dropped Granite 4.2—3B, 8B, and 30B reasoning models trained from scratch on 15T tokens, with thinking/non-thinking modes and a staged RL pipeline that teaches agents to code, search, and use terminals.
Hugging Face's summer 2026 data reveals a seismic shift: Chinese labs now dominate frontier open models, Qwen has become the community's standard, and US contributions have pivoted to hardware vendors.
Allen AI's new embedding exports turn complex Earth observation into surprisingly simple linear algebra. No training data, just dot products and mangrove maps.
Hugging Face and EvalEval just patched the biggest hole in AI benchmarking: scattered, incompatible eval results. Now the same score shows up on model cards *and* links to full reproducibility data.
The full voice loop—speech recognition, Gemma 4 reasoning, and TTS—now runs fast enough to feel natural. Plus: why the P95 latency tail matters more than median response time.
AWS open-sourced Strands Robots, an SDK that exposes LeRobot's stack as composable agent tools. Record demos in sim, push to Hub, run policies, deploy to hardware—all in one agent.
AI2 releases olmo-eval, a modular evaluation framework designed for the iterative reality of training LLMs—not just scoring finished models.
Hugging Face, Meta PyTorch, Nvidia, and a dozen others just formed a committee to govern OpenEnv—the protocol layer trying to make agentic RL training actually interoperable.
JetBrains just released Mellum2, a 12B-parameter MoE model that activates only 2.5B per token. It's not trying to be frontier—it's built for routing, RAG, and agent subtasks where speed matters.
IBM's new 95M parameter embedding model punches way above its weight class with 32K context and true multilingual support—all under permissive Apache 2.0 licensing.
IBM drops the full playbook on Granite 4.1, from data curation to GRPO reinforcement learning. The transparency here is genuinely rare—and the results are competitive.
NVIDIA just open-sourced a state-of-the-art OCR model trained almost entirely on synthetic data. Here's why that matters for the future of vision-language models.