1. DeepSeek previews DeepSeek-V4 model series
DeepSeek released a preview of its DeepSeek-V4 series, featuring the 1.6-trillion-parameter V4-Pro and the 284-billion-parameter V4-Flash. Both models support a one-million-token context window and are available under an MIT license or via a first-party API. The models utilize a new hybrid attention architecture to reduce inference compute and KV cache requirements for long contexts. API pricing is set at $1.74 per million input tokens for V4-Pro and $0.14 for V4-Flash, though early benchmarks indicate high output token usage during reasoning tasks.
2. Sakana AI launches Fugu multi-agent orchestration API beta
Sakana AI introduced the beta version of Sakana Fugu, a multi-agent orchestration system designed to coordinate frontier foundation models. The system dynamically selects optimal agent combinations and roles for tasks like coding and scientific reasoning rather than relying on static rules. Fugu is accessible via an OpenAI-compatible API to allow integration into existing developer workflows. The service is offered in two variants: Fugu Mini for low-latency applications and Fugu Ultra for complex reasoning tasks.
3. Browser Harness library released for LLM browser automation
Developers released Browser Harness, an open-source library that provides LLMs with direct access to the Chrome DevTools Protocol (CDP) for browser automation. Unlike frameworks that restrict models to predefined functions, this tool allows the LLM to write its own click helpers, manage targets, and handle edge cases like cross-origin iframes dynamically. The system relies on a persistent CDP websocket daemon and basic Python helpers that the model can modify on the fly. This approach reduces silent failure modes by giving the model perfect context on how its tools interact with the DOM.
4. AI2 adds custom embedding exports to OlmoEarth Studio
AI2 introduced a feature in OlmoEarth Studio allowing developers to compute and export custom embedding vectors from its open-source OlmoEarth foundation models. Users can configure parameters like area of interest, time range, and imagery sources via the Studio UI or API. The system generates compact numerical representations of Earth-observation data downloaded as lightweight Cloud-Optimized GeoTIFFs (COGs). These embeddings are optimized for downstream applications such as similarity search, change detection, and few-shot mapping.