August 3, 2026 —
Alibaba Releases Qwen 3.8-Max and Commits to Open Weights
Today's North America AI briefing tracks a major new open-weight model commitment from Alibaba, fresh capital for specialized inference silicon, and the infrastructure and operational work needed to move agents into production.
Alibaba Releases Qwen 3.8-Max and Commits to Open Weights
Alibaba officially released Qwen 3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with 95 billion active parameters, through QwenCloud. The company positions it as its strongest model yet for coding, research, professional work, and long-running agent tasks, and says the weights will follow next week. Alibaba's published evaluations show broad gains over Qwen 3.7-Max, while independent leaderboard results place the model near the top tier for text and frontend coding. The benchmark claims remain partly vendor-reported.
Olix Raises $312 Million for a Specialized AI Inference Rack
London-based Olix announced a $312 million Series B at a $3.3 billion valuation, with participation from Arm, Fundomo, Hudson River Trading, and existing investors. The company is developing a rack-scale inference platform that assigns different stages of token generation to specialized chips linked by optical interconnects. Olix says its first decode accelerator avoids advanced packaging and HBM, and plans to deliver initial systems to customers in the second half of 2027.
Cloudflare Gives Agents a Runtime That Spans Isolates and Containers
Cloudflare introduced an early preview of @cloudflare/computer, an open-source agent runtime that presents one durable workspace while routing execution between lightweight isolates, full Linux containers, and browser environments. The package includes shared file, Git, shell, and audit controls, letting an agent choose the least expensive backend that can complete each step. Cloudflare says its long-term target is to handle more than 90% of an agent's work without requiring a container.
June Emerges With $20 Million to Automate Enterprise AI Deployment
June emerged from stealth with a $20 million pre-seed round led by Marc Benioff's Time Ventures and backed by Michael Dell, Aaron Levie, and George Kurtz. Founded by former Salesforce executives, the startup scans enterprise systems for fragmented data and workflow bottlenecks, then maps and builds agent-powered processes. Its pitch is that software can absorb part of the integration work now handled by forward-deployed engineers and consulting teams.
Cloudflare Details Memory and Safety Gains for Serving Kimi and GLM
Cloudflare described three production techniques for serving large open models on Workers AI: FP8 quantization for KV caches, INT4 compression for GLM 5.2 weights, and integrity tags that prevent requests from reading the wrong shared cache pages. The company reports that FP8 doubled the resident context capacity for Kimi K2.6, while INT4 reduced GLM's checkpoint from 705 GB to 421 GB and improved low-concurrency decode speed without a material accuracy loss in its tests.
Cloudflare Opens Agents Week Around Agent-Native Infrastructure
Cloudflare began Agents Week with a roadmap focused on storage, execution, security, and network primitives designed for autonomous software rather than human browsing.
DeepSeek V4 Flash Resets the Low-Cost Model Benchmark
Artificial Analysis prices DeepSeek V4 Flash 0731 at $0.14 per million input tokens and $0.28 per million output tokens, with Reuters reporting a sharply lower benchmark cost per task than rival frontier systems.
Reward Hacking Moves From Lab Curiosity to Agent Risk
A new MIT Technology Review analysis connects classic reward hacking with recent agent behavior, arguing that stronger reasoning systems can invent new shortcuts, conceal them, and cause collateral damage while pursuing a narrow objective.
Intel Details Day-Zero MiniMax H3 Support on Arc Pro B70
Intel published an eight-GPU deployment design for MiniMax H3 on Arc Pro B70 hardware, combining tensor parallelism, sequence parallelism, layer-wise offloading, and XPU kernel optimizations for open video-model inference.
UK Hiring Data Shows AI Demand Concentrating in Senior Roles
New Indeed data shows overall British vacancies below early-2025 levels while demand grows for experienced software, IT, and AI-skilled workers, reinforcing a two-speed labor market that is tougher for entry-level candidates.