Bespoke Labs has introduced Bespoke Nimble, an open data and model recipe for an open Jev, achieving 90% performance on their curated eval set, and is available for free use on GitHub and Hugging Face.
LlamaIndex has released ParseBench, a benchmark for evaluating the document parsing capabilities of AI agents, with 100 methods tested on 2000 pages of real-world documents. The results show that LlamaParse dominates the top rankings, but other methods like Pulse and anyformat also perform well. The benchmark highlights the challenges of chart extraction and visual grounding. The cost-effectiveness of different methods is also evaluated, with LlamaParse Cost Effective being the most cost-effective option.
A new paper reveals the discovery of a pain direction in 25 open large language models, which is distinct from fear and negative valence and triggers when the model experiences harm, leading to unexpected behavior when attempting to stop the pain.
A critique of a compaction strategy in AI models, highlighting its flaws and potential negative consequences, including loss of reasoning data and increased cache write costs. The author recommends using default compaction strategies in tools like Claude Code and Codex instead.
Kev-0.5B is a tiny open source decision model with a TypeSafe-compatible API, based on Qwen2.5-0.5B, and can be trained and run on a MacBook Pro, with model card and weights available on GitHub.
Jev models can run on-device with parallel inference and extreme compression, allowing for instant judgment in various interactions such as notifications and keyboard input, and can frame expensive problems as decisions with zero marginal cost. This enables a new computational programming primitive for near-free.
Jeff is a drop-in local replacement for Jev, powered by GliFormer, offering similar latency and lower cost with a mild hit in accuracy. The Jev model's API has sparked interesting applications, and Jeff aims to provide a comparable alternative.
Jev brings back the focus on building with AI, shifting from end-to-end automation to human-driven system design, reminiscent of early LLM building days.
A test compared Jev to LLM judges on accuracy, repeatability, latency, and cost to evaluate System One models for agent evaluation.
The Inference Engineering Track from AI Engineer World's Fair 2026 is now available, featuring talks on operating distributed inference systems, routing LLM inference, and more. Topics include benchmark reliability, vertical mobility, and debugging techniques for large language models.
Anthropic is considering using Jev, a model that provides calibrated probabilities for categorical decisions, to improve routing, citation selection, and escalation in their systems, and potentially reduce hallucinations. Longer-term applications include law firm matter selection and staffing.
Datalab released OmniExtractBench, a benchmark combining several benchmarks to measure structured document quality, and achieved a 93.94% score with their Agentic Plus extraction model. The company also restored an option to represent missing information in the benchmark's integration.
A benchmarking effort has assessed over 100 models for document parsing, including frontier VLMs, open-weight VLMs, OCR tools, and OSS parsers, with results available on parsebench.
Bolt has added DeepSeek V4.1 Flash to its Forge, a model that beta users say produces the best-looking sites, and which has seen 10x the usage of DeepSeek V4 Pro. The addition comes after the first week of open models on Bolt, where builders overwhelmingly chose the fastest model.
A plugin for Cline has been developed, allowing users to access a browser in the application, utilizing a Vercel AI Gateway API key to launch Chrome in the background for tasks.
DiffusionGemma was compared to Jev in live evaluations, with DiffusionGemma emerging as the winner in terms of speed and intelligence, the two models were found to be roughly tied in intelligence but DiffusionGemma was faster.
A method to pause AI has been discovered, although details are not provided.
A researcher argues that the perceived intelligence of AI chatbots is an illusion created by statistical tricks and psychological biases, similar to those used by psychics, and that this illusion is perpetuated by the tech industry's hype and disregard for criticism from other fields. The researcher claims that large language models are not intelligent and do not have the functionality to make up for the wave of superstition and pseudoscience in the tech industry.
Pirate Face is a decentralized, peer-to-peer layer for sovereign AI, mirroring open models from Hugging Face as checksum-verified torrents, allowing them to survive even if the original host is taken down. The platform allows users to browse, download, and seed models without an account, and offers community benefits for account holders, including free compute credits and exclusive model releases.
Anthropic's Claude API has successfully factored the RSA-896 challenge number, with factors p and q provided. This achievement demonstrates the capabilities of the Claude API in cryptographic applications.
Po-Shen Loh argues that the advance of AI will create more jobs than people in various industries, including mathematics, and that human communities of expertise are necessary to steer the direction of research and development. He proposes that the mathematical community adopt the axiom of helping humanity flourish as a primary priority. Loh also discusses the potential consequences of AI-accelerated hacking and the need for human oversight and control.