News

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati, released Inkling-Small on July 30 — two weeks after the 975B flagship Inkling. The small model has 276B total and 12B active parameters, runs under an Apache 2.0 licence, and scores 40 on the Artificial Analysis Intelligence Index, within one point of the flagship's 41. It is designed for enterprise fine-tuning via the lab's Tinker platform.

DeepSeek released V4-Flash-0731 on July 31, an updated version of their 284B MoE Flash model re-post-trained specifically for agentic and coding performance. TerminalBench 82.7 and NL2Repo 54.2 are the headline numbers. The Artificial Analysis Intelligence Index puts it 10 points above the previous Flash and competitively close to Pro-tier models. MIT-licensed weights and GGUF versions are already available.

Alibaba released Qwen 3.7 Flash on July 27, a cost-optimised multimodal model priced at $0.03 input / $0.13 output per million tokens with a 1M context window. It accepts text, image, and video input and is designed for high-volume agent loops, browser automation, and support tooling. It is the flash-tier counterpart to the Qwen 3.7 Max, Alibaba's flagship agentic model released in May.

KwaiKAT, the AI team at Kuaishou, released KAT-Coder-V2.5 — an agentic coding model trained end-to-end on 100,000+ real repositories that autonomously navigates codebases, implements multi-file changes, and runs its own tests. A 35B open-weight MoE dev version shipped to Hugging Face alongside the commercial Pro release.

Google updated the Managed Agents feature in the Gemini Interactions API with two production-readiness upgrades: asynchronous background execution (jobs that survive dropped connections) and native remote MCP server support (no proxy middleware required). Gemini 3.6 Flash is now the default model powering both.

Microsoft announced MAI-Cyber-1-Flash, its first in-house AI model built specifically for finding and managing software vulnerabilities. Paired with Project Perception, its new agentic security system, it scored 96% on the CyberGym benchmark at roughly half the cost of its previous configuration.