Xiaomi open-sources MiMo-V2.6, a 1.02T-parameter model family
Xiaomi released and open-sourced its MiMo-V2.6 model family on September 21–22. The flagship MiMo-V2.6-Pro is a sparse mixture-of-experts model with 1.02 trillion total parameters and 42 billion active per token, running a 1M-token context window. Weights for Pro, Flash and a 9B distillation are on Hugging Face under an MIT license. Artificial Analysis scored the Pro model at 46 on its Intelligence Index, first among open-weight systems. Xiaomi streamed a six-day reinforcement learning run that cost about $2.62 million for Pro, and shipped the training framework with more than 7,000 RL environments alongside the weights. API pricing stays at V2.5 levels: $0.435 per million input tokens and $0.87 per million output tokens.
Source: Xiaomi, Silicon Report
Anthropic ships Claude Opus 5.5 with lower running costs
Anthropic introduced Claude Opus 5.5 on September 22. Standard API pricing is $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. The company reports roughly 40% lower costs than Opus 5 on typical workloads at default settings, a vendor-measured figure. Anthropic also says the model sets its own reasoning depth, so fixed "think step by step" prompting tends to add latency without improving output.
Source: Anthropic, via CodeMingle
OpenAI adds GPT-6 Sol and Luna at half the promo price
GPT-6 Sol and Luna went live on September 24, filling out the GPT-6 line below Astra. Sol costs $2 per million input tokens and $10 per million output tokens; Luna runs $0.10 and $0.50. Both prices sit 50% under GPT-5.6 promo rates. A quieter change matters for agent builders: switching reasoning effort or tools no longer invalidates the prompt cache, which removes a full-price context refill on agent stacks that raise effort only for hard steps.
Source: OpenAI, via The AI Wire
Alibaba details its chip and model roadmap at Apsara
At its Apsara Conference, reported September 26–27, Alibaba's T-Head unveiled the Zhenwu V900 accelerator with 216GB of memory per card and 1.2TB/s of chip-to-chip bandwidth, a 3x uplift over the M890 by the company's numbers. Mass production is slated for Q1 2027, and single deployments can scale to 500,000 cards. Alibaba also said Qwen 4 is in training while Qwen 4.5 and Qwen 5 target 5–10 trillion parameters, and it set a goal of more than 20GW of cloud capacity by 2032.
Source: The Neuron, AGI Hunt
Anthropic commits $11.6B to Akamai cloud
Anthropic signed a seven-year, $11.6 billion commitment to Akamai's cloud infrastructure, announced September 27. It is the largest contract in Akamai's history. Warrants for up to a 5% equity stake are tied to spending milestones, with terms that could lift total spend toward $20 billion. Most of the capacity is general-purpose CPU rather than GPU, reflecting the data-retrieval and code-execution work that agentic workloads generate.
Source: AI Breaking Wire
NVIDIA makes ROS development agent-friendly
NVIDIA released Isaac ROS 5.0 at the ROSCon conference in Toronto on September 22. The release adds support for ROS 2 Lyrical and Ubuntu 24.04, reusable skills for setup and manipulation that developers and AI agents can both invoke, and an updated FoundationPose inference library for pose estimation and tracking that NVIDIA says runs up to 5.5x faster. NVIDIA also contributed a standard data-handling interface, rosidl::buffer, to the ROS 2 Lyrical long-term-support release through the Open Source Robotics Alliance. The stack is free and open source, and it deploys across Jetson hardware from the Orin Nano to Jetson Thor.
Source: NVIDIA
Alphabet's Intrinsic open-sources Intrinsic Core
Intrinsic, the Alphabet robotics software unit, released Intrinsic Core as open source under Apache 2.0, reported September 24. The suite ships preconfigured runtime services for industrial robotics, including an Open Machine Tending Solution reference application for CNC machine tending that integrates NVIDIA FoundationPose for part detection and pose estimation out of the box. The design lets robots handle parts dynamically with fewer rigid fixtures.
Source: AI Impact Hub, NVIDIA
Epson launches its first collaborative six-axis robot
Epson Robots announced the AX6 on September 22, its first collaborative 6-axis robot. The arm uses a carbon fiber structure, carries an IP54 rating and meets ISO Class 5 (ISO 14644-1) cleanroom standards, and supports 100–240V AC and 48V DC power. Epson pairs it with no-code software and an AX Portal environment, with a Python option for advanced applications.
Source: Epson Robots, via Unite.AI
A machine-tool cluster in Taizhou pushes AI-driven machining
Xinhua reported on September 22 that the Zhejiang High-end CNC Machine Tool Technology Innovation Center has rolled out an AI-driven, fully direct-drive CNC machine. Inspection by the National Machine Tool Quality Supervision and Inspection Center put overall performance 30–60% above national standards, with machining efficiency up as much as 58% in some working conditions. The direct-drive design routes motor power straight to the tool end, and one machine under commissioning in Wenling holds 4-micron accuracy through 72 hours of continuous operation. The center has also built a design resource library of 1,500 machine modules.
Source: Xinhua
Ready to start your HMI project?
Our engineers help you choose the right HMI hardware and software for your project.