AI
Artificial intelligence developments, regulation, and industry
DeepSeek upgrades open V4.1-Flash model and launches Search APIs
- DeepSeek V4.1-Flash-1017 and Search API: DeepSeek released an upgraded checkpoint of its 552-billion-parameter Mixture-of-Experts model under an MIT licence on Hugging Face, achieving 77.0 on DeepSWE v1.1 and 42.3 on Terminal-Bench 4.0 while shortening output length and task runtime at unchanged API prices. Alongside the model weights, the lab opened its in-house Search and Fetch APIs to developers at $0.75 per 1,000 calls, with server-side tool calls inside coding harnesses remaining free until 31 October.
- Jais 3 Arabic Agentic and Speech Fleet: Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence, Inception42, and Cerebras Systems launched Jais 3, a family of six open-weight Arabic models ranging from 900 million to 375 billion parameters built on the K2 Horizon foundation. Supporting a 512,000-token context window (128,000 on the smallest model) and paired with the Jais Speech recognition system, the stack handles Modern Standard Arabic, 19 national dialects including North African Arabic, and mid-sentence code-switching into French and English for local enterprise and public-sector workflows.
- Reflection Beam 1.0: US lab Reflection AI released its first open-weight model, Beam 1.0, under an Apache 2.0 licence, featuring 501 billion total parameters, 23 billion active parameters per token, and a 1-million-token context window. Built for token-efficient reasoning and priced at $0.25 per million input tokens and $1.00 per million output tokens on API providers, Beam 1.0 scored 30 on the Artificial Analysis Intelligence Index and 48.5% on AutomationBench-AA while using roughly half the output tokens of comparable open models.
- Tencent Hunyuan Hy Memory 2.0: Tencent open-sourced a three-layer agent memory stack anchored by Hy-Memory-30B-A3B, an Apache 2.0-licensed Mixture-of-Experts model that activates 3 billion parameters per token specifically to manage long-term memory writing across Mem0, MemOS, and Hy-Memory frameworks. Paired with the local MemHub application, the specialised writer replaces costly general-purpose frontier models in deciding what project context and code state agents retain across sessions.