join.dev
AI news · 10 stories
Reflection releases Beam weights and its agent tooling stack · Meta releases Muse Spark Open for multimodal agent work · NVIDIA releases Nemotron Super VL and Bolt models
Reflection released Beam’s weights under Apache 2.0 on October 16, alongside its API and open agent integrations. The release moves beyond the earlier preview, giving developers tools to self-host, customize and deploy the model.
Read at Reflection
Meta released Muse Spark Open, an 816-billion-parameter multimodal model with a one-million-token context window, under its Muse Model License. Hosted APIs and downloadable weights broaden deployment options, including an optimized setup using two M5 Ultra Mac Studios.
Read at Meta
NVIDIA released Nemotron 3.5 Super VL for multimodal enterprise agents and Bolt 74B-A9B for local workloads fitting in 24GB of memory. The paired releases extend customizable models from server deployments to desktop hardware; the smaller Bolt E4B is scheduled for October 23.
Read at NVIDIA
Mistral introduced Timestral, a family of time-series forecasting models ranging from 5 million to 209 million parameters that can run on CPUs. Probabilistic forecasts and support for related variables give teams a compact option for planning and monitoring industrial data.
Read at Mistral
Elorian introduced Glimpse on October 16 as a research preview that revisits visual evidence while reasoning about images and motion. API access lets developers explore tasks involving object tracking, changing viewpoints and physical consequences.
Read at Elorian
Anthropic reports that Claude improved published quantum-circuit designs across eight tasks using an automated evaluator to check and score proposals. The results suggest AI-assisted algorithm design could reduce the hardware resources needed for useful quantum computations.
Read at Anthropic
Anthropic reports that an internal Claude research model traced NIST elliptic-curve seeds to the lrand48 random-number generator initialized with Unix timestamps. The finding addresses a longstanding provenance question and does not constitute a cryptographic break or a newly discovered weakness.
Read at Anthropic
Anthropic made Code Review generally available to Claude Team and Enterprise organizations, with automatic reviews on new pull requests. Repository-specific review depth and faster turnaround aim to ease the review bottleneck as coding agents increase output.
Read at Anthropic
Attie announced a site builder for small social experiences such as question-and-answer pages, guestbooks, surveys and games. The expansion gives users a way to create interactive projects and share them through Bluesky.
Read at Bluesky
Anthropic introduced public-beta OpenTelemetry traces for Claude Managed Agents, the Messages API and Claude SDKs. Exporting model calls, tool use, latency and token information into existing observability systems helps teams investigate agent behavior and performance.
Read at Anthropic