
Sapiens AI's Agnes 2.5 Pro Beta Triples Agentic Performance at Bargain Prices
Singapore lab Sapiens AI pushes its Agnes 2.5 Pro model to 49 on Artificial Analysis Intelligence Index through agentic gains, but at roughly double the token cost.
20 stories

Singapore lab Sapiens AI pushes its Agnes 2.5 Pro model to 49 on Artificial Analysis Intelligence Index through agentic gains, but at roughly double the token cost.

Chroma unveils Fission, a concurrency protocol for agent swarms that skips rollbacks to preserve expensive reasoning tokens when writes collide.
Inco AI's DFlash 2 drafter for GLM-5.3-Flash now ships under Apache 2.0, delivering near 3x speculative decoding speedups with lossless output.

A community fine-tune of Qwen3.8-27B strips refusals, fixes the resulting weight damage, and ships in four formats tuned for local inference.

Anthropic opens Claude Team seats to 10,000 academic researchers with free Standard access and $15 Premium seats, an 80% cut off list price for a year.

Google Research unveils an autonomous AI system that turns natural-language questions about food security, disease, and climate risk into trained geospatial models in minutes.

Anthropic unveils a shared specification that lets AI agents discover and safely operate lab and factory hardware, cutting weeks of integration to hours.

A UIUC and Bridgewater team fine-tuned Kimi-K2.6 on Tinker to become the first text-to-SQL model to beat human accuracy.

Google's video model gets scene extension to 40 seconds, keyframe control, 360p drafting, 4K upscaling, and video reference inputs.

Krea launches MiniMax H3 Max, a speed-tuned variant of the Hailuo 3 video model that renders 15 second clips in roughly 5 seconds.

Brett Adcock's stealth AI startup lands gigawatt-scale Vera Rubin capacity, joining Thinking Machines and xAI in NVIDIA's biggest compute alliances.

A new keypoint detector skips deblurring entirely, learning directly from blurred images through self-supervision and beating supervised baselines on matching and localization.

A local Python CLI parses session logs from Claude Code, Codex, and Gemini CLI to tally token usage and spend by model, project, and day.

Cohere's new 2.3B-parameter vision model turns PDFs into agent-ready Markdown at $1.50 per 1,000 pages, undercutting Mistral by 63%.

Google DeepMind's new cryptographic testing setup lets outside auditors evaluate Gemini without ever seeing model weights or leaking their prompts.

QoderAI's Better Harness plugs into Claude Code, Codex, Cursor and Copilot to audit how your coding agents actually work, not just what they ship.

A community developer shipped a 54-node ComfyUI extension for MiniMax H3 with same-day tooling for Alibaba PAI's new 8-step acceleration LoRAs.

A community project squeezes Qwen3.8-27B onto a 24GB gaming card with vLLM, hitting 417 tok/s batched or 82 tok/s single-user at 150k context.
A 1.27 GB Q4_K_M build of Qwen3.8-2B with its refusal direction surgically removed, aimed at CPU-only laptops and phones.
An independent quantizer squeezed Z.ai's 320B-parameter GLM-5.3-Flash into a 4-bit checkpoint that runs multimodal inference on two workstation GPUs.