AI — Page 17

Blackroot's new guide translates a January architecture that trades some MoE experts for hashed N-gram embeddings.

The same-day follow-up puts Qwen's new model on QwenCloud with a default 1-million-token context and OpenAI and Anthropic API compatibility.

The former Meta researchers behind Isaac 0.5 are betting one model can handle perception, reasoning and control without a stack of specialized systems.

Customers get Reuters journalism dating to 1987 in five languages, while pricing and permitted model-training uses remain undisclosed.

The Beijing lab told Bloomberg it built the anonymous coding model and planned to release its weights on August 26th.

The 27-billion-parameter open-weight model ranks seventh overall in a human-preference test of visual web development, a useful result that says little about production code quality.

ModelBest's Pinea Pi pairs local MiniCPM models with Intel or Nvidia compute and integrated sensors ahead of a planned Kickstarter campaign.

Thinking Machines Lab will subsidize outside safety work on Inkling, although it has not disclosed the program's budget or number of awards.

Founder Yan Junjie's efficiency thesis gets a tidy agent demo, though the visible test does not establish a repeatable cost.

The campaign used ChatGPT for posts and channel logos while an Israel-branded website republished academic work and praised Russia.

Premium costs five times the Standard rate and includes 5x usage, while letting workspace owners mix both seat types across a team.

Brad Wardell's Clairvoyance 0.85 lets colleagues work with the same agents while the host computer retains the session, files and local execution.

OpenAI reports gains over Nvidia GB200 and GB300 systems in company-run tests, with production deployment planned by the end of 2026.

Figure says 44,000 weekly creators have uploaded 16M videos, and it plans to spend more than $1B on data and compute over the next year.

The first-party accelerator beat GB200 and GB300 systems in OpenAI-run tests, with deployment planned inside OpenAI's infrastructure by December.

The model learns unseen tasks lasting up to 10 minutes without fine-tuning, according to tests published by the robotics developer.

The anonymous coding model matched GLM-5.2's tokenizer and heavily suppressed seven topics tied to Xi, party legitimacy and domestic protest.

IBM's 470M-parameter English models trade translation and keyword biasing for lower latency, smaller memory needs and edge deployment.

Anandkumar and Benedikt Jenik kept building Accelerated Understanding after declining a proposed Prometheus package; the company says its physics-focused model handled 5 trillion pieces of data in one prompt.

The Laude-MIT collaboration runs a self-directed thought loop in less than 10,000 lines of Bash, with continuous token costs and familiar security problems.

Its 255,068-skill comparison shows multilingual growth, while lexical discovery remains a likely constraint on cross-language reuse.

Reverse engineer Xusheng Li found that Paint and Photos alter output pixels with a GUID untouched by the visible-watermark setting.

The new index covers 24GB local deployment, ComfyUI and multi-GPU serving, three weeks after MiniMax released H3's weights.

Internal documents point to a late August or early September rollout, followed by Meta's Watermelon model in October.