Most visible doc retrievers in manufacturing at this time are hand-me-downs. ColPali and the fashions that adopted it take a generative vision-language mannequin and repurpose it as an encoder. The consequence nonetheless carries a individually pretrained imaginative and prescient tower and a causal decoder that by no means generates a token. That is parameter and…
A staff of researchers from UC Berkeley have launched CUA-Lite, an open platform for computer-use brokers (CUAs). The argument behind it’s infrastructural moderately than model-centric: coaching and benchmarking a CUA requires 4 items: brokers, environments, traces, and a framework to judge and prepare them and all 4 are at present fragmented throughout separate repositories with…
GitHub has launched Project HydraFusion, a analysis preview that stops treating mannequin selection as a one-time setting. Instead of routing your immediate to a single mannequin, HydraFusion builds an execution plan per request. It can draft with one mannequin, have a second mannequin critique the draft, or escalate to a stronger mannequin when a high…
The laborious a part of operating an open-weights mannequin regionally was by no means the mannequin. It was every thing earlier than it, studying VRAM specs, guessing which quantization matches, setting context size and GPU layer counts, then discovering out at load time that the file is three gigabytes too massive. Nous Research has collapsed…
Video has been the costliest modality to motive over. A Gemini mannequin handed a 90-minute lecture has, till now, ingested the entire thing at a set one body per second, whether or not the query was ‘summarize this’ or ‘what time does the speaker change to the pricing slide?’ That single-pass design forces a nasty…
Every humanoid startup, autonomous-vehicle refugee workforce, and frontier AI lab is fishing in the identical tiny pond, and paying startup-lottery cash to do it. If you construct software program for a dwelling and have puzzled whether or not that talent carries right into a warehouse filled with strolling robots, the reply this yr is: greater…
Today, OpenAI launched GPT-6 Astra. The firm calls it its most clever and aligned mannequin, and positions it primarily as a computer-use system relatively than a chat mannequin. The pitch is that Astra operates software program the best way a individual does, throughout browsers, spreadsheets, desktop functions and terminals, and finishes multi-step jobs as an…
Most groups constructing a procuring assistant or agent rebuild the identical scaffolding: an agent loop, a device layer over the catalog, an approval gate, and an eval suite. Anthropic has now launched that scaffolding as code. This week, they printed anthropics/commerce-agents, a reference blueprint containing a procuring agent and a service provider agent, together with…
This week, Meta Superintelligence Labs launched Muse Spark 1.3. It is the fourth Muse Spark launch in 5 months, and the goal is long-horizon agentic and coding work somewhat than single-turn technology. The framing in Meta’s submit is usability: sustaining an extended thread, collaborating with the person, and understanding when it’s caught. Is it deployable?…
Perplexity has open sourced Lily, the native inference engine behind Hybrid Compute in Perplexity Computer. It is a single-process runtime: a Rust layer masses the checkpoint and drives the era loop, an OpenAI-compatible chat-completions API streams tokens, and hand-written Metal kernels execute the mannequin. Neither PyTorch nor MLX sits within the execution path. Lily is…