Models
uni-robot-agent: What Qwen3.8-Flash-Next's 31.39 Really Beats
Qwen3.8-Flash-Next tops GPT-6 Astra on RoboDojo, 31.39 to 28.97, but only against Astra's baseline harness. On the same harness, Astra scores 51.90.
Models
New and updated AI models, benchmarks and open-weight releases, explained in plain language.
Models
LlamaIndex OCRia-1 Inference: What the 10x Speedup Measures
LlamaIndex says OCRia-1 runs nearly 10x faster on one B200. The technical report shows that's single-stream speed; pages per second rise about 4.7x.
Models
Google Suncatcher Gemma 4 Orbit Test: What the Haiku Proves
Google's Suncatcher satellite ran Gemma 4 on a TPU in orbit. One prompt, one of four chips on at a time: what the haiku proves and what is still untested.
Models
Xiaomi-OCR-0: Why a 95.24 Real5 Score Became 92.73
Xiaomi says its OCR model scores 95.24 on Real5; the benchmark lists 92.73 for the same outputs. A public thread traces the gap to the scoring code version.
Products
Cloudflare Zero Wire: What the AI Prompt and WAF Chart Show
Cloudflare's Zero Wire is a curated feed of critical CVEs. The AI part is a prompt you run yourself, and the WAF chart counts rule matches, not hacked sites.
Business
Anthropic's Amended S-1: What the $895.4B Commitment Includes
Anthropic's amended S-1 shows $17.0B in Q3 revenue and $895.4B in compute commitments, but about $213.8B of the total can be terminated or is conditional.
Products
AMD's Agent Profiling Tutorial for Instinct GPUs: What It Proves
AMD's ROCm tutorial profiles an AI agent on Instinct GPUs and shows a slow tool going from 83.8 s to 7.0 s. Those are example values, not measurements.
Products
NVIDIA AICR for Dynamo: What 'Validated' Means on Kubernetes
NVIDIA's AICR recipe pins 17 Helm charts for Dynamo 1.4.2 on Kubernetes. What "validated" covers, what it leaves to you, and one alpha release in the stack.
Products
Obsidian 1.15 Adds Pele Diagrams and iOS Voice Capture
Obsidian 1.15 swaps in the Pele Mermaid renderer, adds iOS 26 voice capture and brings PDF export to mobile. What changes for plugin authors and iPhone users.
Products
Claude Message Threads Cut Agent Latency by 35%
Claude message threads, now in beta, cut time to first token about 35% in 500k-1M token sessions. Token costs stay the same, and threads expire in 5 minutes.
Products
DeepSeek Harness Super-agent: An Always-On AI Agent Design
The DeepSeek Harness team wants one always-on AI agent that juggles many tasks across chat and GitHub. It's a proposal, and early tests are small and simulated.