Public story · 2026-03-08 · source-backed

AI2 Olmo Hybrid 7B — 2x Training Efficiency via Hybrid Architecture.

Issue

2026-03-08

Confidence

source-backed

Sources

Redaction

passed

Story

Fully open model using 3:1 DeltaNet-to-attention that matches Olmo 3 with 49% fewer tokens. Novel theoretical proof that hybrids are strictly more powerful than either transformers or RNNs alone. Trained on 512 GPUs. Full weights, checkpoints, and code released. AI2 Blog

OLMo Hybrid 7B: 2x Data Efficiency Proves the Transformer-Only Era May Be Ending.
AI2's fully open 7B model replaces 75% of attention layers with Gated DeltaNet linear recurrence, matching OLMo 3 accuracy on MMLU with 49% fewer tokens. Nathan Lambert's analysis highlights the theoretical contribution: hybrid models formally solve problems that neither pure...
Anthropic Gets 220,000 GPUs. Your Claude Code Limits Just Doubled.
This one hit my inbox and I had to read it twice. Anthropic announced a partnership with SpaceXAI for the entire Colossus 1 data center in Memphis. 220,000 NVIDIA GPUs. 300+ megawatts. That's the largest single compute acquisition by any AI lab. Full stop. But the part that ma...
Builder Skills
1. Defend Against Vibeware DDoD — Behavioral process monitoring for AI-generated polyglot malware. Monitor Living Off Trusted Services patterns. Bitdefender 2. OS-Level Agent Sandbox — Kernel-level sandboxing (Seatbelt/Bubblewrap/AppContainer) for agentic workflows. Applicatio...
Soofi S: Germany's sovereign 30B-A3B MoE trained on 27T tokens, entirely on Deutsche Telekom infrastructure.
Soofi S (Sovereign Open Source Foundation Models, IPCEI-CIS/8ra-funded) hit the HN front page with a 30B mixture-of-experts activating 3B parameters per token, pretrained on ~27 trillion tokens with deliberate German weighting. The paper (arXiv:2607.09424) claims it matches de...
A Hidden HTML Comment in a Pull Request Owns Your Review Agent, and Microsoft Hasn't Patched It
Someone opens a PR against your repo. The description looks normal in the browser. Buried in it is . Invisible in the Azure DevOps web UI. Fully visible to your review ag...
ZGCM-1 is a fully open 7B whose training cluster was operated by agent swarms.
arXiv 2609.13356, 227 HuggingFace upvotes, releases a 7B dense model trained from scratch on the premise that small models can't memorize the web but can trade parametric capacity for deliberate thinking plus external tools. Fully open recipe: interleaved gated sliding-window...

Source trail

AI2 Blog More from allenai.org

Entities

AI2 Blog AI2 Olmo Hybrid DeltaNet Full Fully GPUs Hybrid Architecture Novel Olmo RNNs Trained Training Efficiency

Provenance

Canonical issue: Ramsay Research Agent — 2026-03-08
AI generated: no
Story unit: 2026-03-08-ai2-olmo-hybrid-7b-2x-training-efficiency-via-hybrid-architecture
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

AI2 Olmo Hybrid 7B — 2x Training Efficiency via Hybrid Architecture.

Story

Related stories

OLMo Hybrid 7B: 2x Data Efficiency Proves the Transformer-Only Era May Be Ending.

Anthropic Gets 220,000 GPUs. Your Claude Code Limits Just Doubled.

Builder Skills

Soofi S: Germany's sovereign 30B-A3B MoE trained on 27T tokens, entirely on Deutsche Telekom infrastructure.

A Hidden HTML Comment in a Pull Request Owns Your Review Agent, and Microsoft Hasn't Patched It

ZGCM-1 is a fully open 7B whose training cluster was operated by agent swarms.

Source trail

Entities

Provenance

AI2 Olmo Hybrid 7B — 2x Training Efficiency via Hybrid Architecture.

Story

Related stories

OLMo Hybrid 7B: 2x Data Efficiency Proves the Transformer-Only Era May Be Ending.

Anthropic Gets 220,000 GPUs. Your Claude Code Limits Just Doubled.

Builder Skills

Soofi S: Germany's sovereign 30B-A3B MoE trained on 27T tokens, entirely on Deutsche Telekom infrastructure.

A Hidden HTML Comment in a Pull Request Owns Your Review Agent, and Microsoft Hasn't Patched It

ZGCM-1 is a fully open 7B whose training cluster was operated by agent swarms.

Source trail

Entities

Provenance

Story

Related stories

OLMo Hybrid 7B: 2x Data Efficiency Proves the Transformer-Only Era May Be Ending.

Anthropic Gets 220,000 GPUs. Your Claude Code Limits Just Doubled.

Builder Skills

Soofi S: Germany's sovereign 30B-A3B MoE trained on 27T tokens, entirely on Deutsche Telekom infrastructure.

A Hidden HTML Comment in a Pull Request Owns Your Review Agent, and Microsoft Hasn't Patched It

ZGCM-1 is a fully open 7B whose training cluster was operated by agent swarms.

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance

Story

Related stories

OLMo Hybrid 7B: 2x Data Efficiency Proves the Transformer-Only Era May Be Ending.

Anthropic Gets 220,000 GPUs. Your Claude Code Limits Just Doubled.

Builder Skills

Soofi S: Germany's sovereign 30B-A3B MoE trained on 27T tokens, entirely on Deutsche Telekom infrastructure.

A Hidden HTML Comment in a Pull Request Owns Your Review Agent, and Microsoft Hasn't Patched It

ZGCM-1 is a fully open 7B whose training cluster was operated by agent swarms.

Tomorrow's issue, in your inbox.

Source trail

Entities

Provenance