Reflection previews open-weight coding model Beam
Reflection says Beam is a 501B-parameter sparse MoE model with 23B active parameters, focused on coding, reasoning, and agents. A select group can access an early version; the company says weights, documentation, and developer artifacts will follow later this month under Apache 2.0. Reflection reports competitive coding results and 3–4× less inference compute than GLM-5.2 on advanced reasoning benchmarks; these are vendor claims.
Join the early-access waitlist if you want to evaluate it; plan for self-hosting once the promised weights arrive.
Cohere upgrades North with agents and governance
North 2 adds redesigned orchestration for multi-step agents, reusable skills, shared libraries, memory, app creation, and workflow automations. Cohere lists connectors including GitHub, Jira, Slack, SharePoint, and Notion. It supports cloud, VPC, hybrid, on-premises, and air-gapped deployments; North Admin adds model controls, user quotas, rate limits, and spend caps. Cohere directs organizations to contact sales.
For enterprise agent deployments, check whether its deployment options, connectors, and token controls fit your governance needs.
Mistral announces a Large 4 preview
Mistral executive Guillaume Lample announced a preview of Large 4, described as a natively multimodal, 1T-parameter model with 49B active parameters. A social post reports Mistral comparison scores of 61.7% on DeepSWE v1.1 and 50 on Artificial Analysis Cyber Index; these are reported results, not independently verified here. The capture gives no API, weight-access, or pricing details.
Treat the benchmark numbers as preliminary; wait for access and evaluation details before choosing it for coding or security workloads.