Artificial Curiosity Labs
Go back

AI News Digest

AI News Digest — 2026-09-11

Top Stories

DeepSeek Ships V4.1 Flash GA With Causal-Encoder-Decoder MoE as V4 Pro Retires (Pandaily)

DeepSeek’s general availability release of V4.1 Flash, a 552-billion-parameter mixture-of-experts model with native vision capabilities, consolidates its platform by retiring the V4 Pro tier and streamlining pricing—signaling a maturation of frontier-model deployment where multimodal performance and cost efficiency converge in a single offering.

Mistral AI Raises $3.5B At $24B Valuation In Another Record European AI Round (Crunchbase News)

Mistral AI’s Samsung-led $3.5 billion Series D nearly doubles its valuation past $24 billion, marking the largest venture round ever for a European AI startup and solidifying its position as the region’s leading frontier-model builder amid intensifying global competition for large language model dominance.

Maven Robotics wants to steal your robot deployment deal (TechCrunch)

Maven Robotics’ stealth exit with a $100 million Series A and active deployments highlights intensifying investor conviction in physical AI, as capital floods into the operational scaling of robotic systems beyond lab prototypes.

OpenAI puts Pro subscriptions on hold due to Astra demand (techcrunch.com)

OpenAI has paused new Pro subscriptions to manage overwhelming demand for its Astra model, a sign that the surge in advanced AI usage is exceeding the company’s available serving capacity. The temporary halt underscores how even well-capitalized AI firms are confronting infrastructure bottlenecks as they try to scale next-generation models quickly enough.

Chipmaker Positron nabs $875M to speed up inference with consumer-grade memory (siliconangle.com)

Positron’s $875 million raise signals a push to accelerate inference using consumer-grade memory instead of costly HBM, a move that could fundamentally alter the cost structure of large-scale model serving.

Infra & Compute

What moved

China targets 9,800 EFLOPS of intelligent computing capacity by 2030 (TechNode)

China’s new 15th Five-Year Plan target of 9,800 EFLOPS of intelligent computing capacity by 2030, paired with a planned RMB 3.8 trillion investment in information infrastructure, reveals a coordinated state push to massively scale national AI compute supply and build a self-reliant digital backbone.

Huawei showcases a 7.2Tbps near-package optics module for AI infrastructure (TechNode)

Huawei unveiled what it describes as the industry’s first 7.2Tbps near-package optics module, targeting reduced signal loss and lower power consumption for interconnects inside AI data centers.

Cambricon Day-0 Adapts DeepSeek-V4.1-Flash on vLLM Stack (Pandaily)

Cambricon’s Day-0 support for DeepSeek-V4.1-Flash on the vLLM stack signals that Chinese AI accelerators are now shipping day-one software compatibility for frontier open-weight models — a cadence once reserved for NVIDIA — and narrows the practical barrier for enterprises looking to run high-performance inference outside the CUDA ecosystem.

Also tracking

Foundation Models

What moved

OpenBMB Releases MiniCPM5-2B as Open On-Device SOTA Under 4B (Pandaily)

OpenBMB’s MiniCPM5-2B, a permissively licensed 2.5-billion-parameter model, sets a new open-source state of the art for on-device architectures under 4B parameters, combining long-context handling with performance that surpasses some 4B baselines, thereby bolstering the tier of efficient models suitable for edge deployment.

DeepSeek formally launches V4.1 Flash, routes V4 Pro requests to Flash (TechNode)

DeepSeek’s formal rollout of V4.1 Flash and its decision to redirect all V4 Pro traffic to the new model underscore a rapid efficiency gain, with Flash reportedly outperforming Pro in performance, cost, and speed — a follow-through that validates earlier reports of its internal superiority.

JD.com Open-Sources JoyAI-EchoWM Interactive Audiovisual World Model at JDD (Pandaily)

JD.com’s open-source JoyAI-EchoWM, an interactive audiovisual world model with camera-intent control debuted at its JDD conference, adds a publicly accessible building block to the accelerating world-model race, directly enabling simulation and embodied-use-case development through controllable audio-visual generation.

Also tracking

Agents & Tooling

What moved

Alipay to launch an AI wallet agent for trusted AI payments (TechNode)

Alipay’s upcoming AI wallet agent—integrating Vibe Pay, Skill Pay, and Machine Pay—marks a tangible shift toward mainstream agentic payments, where machines autonomously execute transactions within a framework of trusted AI.

Extreme Networks’ Agent ONE Coworker moves AI networking from dashboards to answers (SiliconANGLE)

Extreme Networks’ Agent ONE Coworker exemplifies the agentification wave hitting enterprise IT, transforming network operations from passive dashboards into an interactive conversational model that delivers direct answers instead of raw data, cutting the interpretation burden and accelerating response times.

Salesforce introduces Enterprise AI Harness, AI Control Plane (siliconangle.com)

Salesforce’s preview of enterprise-grade agent-building and governance tooling—branded as Enterprise AI Harness and AI Control Plane—signals that major SaaS platforms are now racing to supply the management layer required to orchestrate fleets of AI agents, a development that follows up on initial reports from August 5.

Also tracking

Robotics & Physical AI

What moved

ACE Robotics and NTU Open-Source Puffin-World Multimodal World Model (Pandaily)

ACE Robotics and NTU have released Puffin-World as open source, providing a unified multimodal world model that captures physics, geometry, and appearance with state-of-the-art pose accuracy; this matters because it gives robotics and embodied-AI teams a reusable foundation for simulation, planning, and perception instead of forcing each lab to build its own world model from scratch.

Robots Are Learning to Feel (IEEE Spectrum)

The growing availability of tactile data is unlocking dexterous manipulation, a critical frontier for general-purpose robotics.

Skild AI Taps NVIDIA Physical AI to Teach Robots New Tasks From a Single Video (blogs.nvidia.com)

Skild AI’s S1 foundation model can learn complex, multi-step robotic tasks from a single demonstration video, signaling a shift toward more practical and scalable robot training by drastically reducing the data and programming typically required.

Also tracking

Policy & Safety / Other

What moved

Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek (techcrunch.com)

Anthropic has detailed specific distillation campaigns from Alibaba, Moonshot AI, and DeepSeek, alleging that the labs persistently extracted its models’ outputs to train competing systems—an update that intensifies the battle over model intellectual property, sharpens the policy debate around extraction safeguards, and heightens competitive pressure on AI labs to protect proprietary data.

Panic builds over bankrupt Spirit’s looming data sale to Google (Ars Technica)

The potential sale of Spirit’s user data to Google through bankruptcy proceedings threatens to set a precedent where companies can sidestep their own privacy commitments and sell personal information for AI training without user consent, simply by declaring insolvency.

OpenAI Wants to Know if an AI Industry Slowdown Would Even Be Legal (wired.com)

OpenAI’s inquiry into the legality of pausing AI development highlights an emerging tension where antitrust law could be weaponized to prevent coordinated safety measures, forcing the industry to navigate between reckless acceleration and potential collusion claims.

Also tracking

Applications

What moved

India’s Pocket FM doubles revenue run rate to $500M as AI powers 93% of audio content (techcrunch.com)

Pocket FM’s doubling of its revenue run rate to $500 million, with 93% of audio content now produced by AI, demonstrates the seismic impact of cost efficiency in media: AI-driven production costs have been slashed by roughly 80 times, enabling a business model that rapidly scales content libraries and revenue while keeping margins healthy.

Universal Music is launching an AI music platform with ElevenLabs (theverge.com)

Universal Music’s partnership with ElevenLabs on a licensed AI music platform signals a strategic pivot for a major label from suing AI companies to building commercial products, enabling authorized remixing that embeds generative tools directly within the industry’s business model.

Healthcare AI’s next test is integration (MIT Technology Review)

The transition from building powerful clinical AI models to embedding them into hospital workflows marks a critical juncture for the sector. Major technology firms are now delivering systems that can parse complex patient records, yet the true obstacle lies in harmonizing these tools with existing electronic health records, clinician routines, and regulatory frameworks—where seamless deployment matters more than raw performance.

Also tracking

Funding & Capital

What moved

Ayar Labs bags $150M in additional late-stage funding to help make bigger AI chip clusters (SiliconANGLE)

Ayar Labs’ $150 million in additional late-stage funding underscores the importance of optical interconnect technology for scaling AI chip clusters beyond today’s bandwidth limits, signaling that capital is flowing to the connectivity layer needed to support larger AI systems.

Maven Robotics wants to steal your robot deployment deal (TechCrunch)

Maven Robotics surfaced from stealth with a $100 million Series A and active deployment contracts, a signal that venture capital is betting heavily on startups that can move physical AI from prototype to production right now — not in some distant future, but in operating facilities today.

Chipmaker Positron nabs $875M to speed up inference with consumer-grade memory (siliconangle.com)

Positron’s $875 million funding round signals a structural shift in AI hardware away from the HBM supply constraints and cost premiums imposed by a single supplier; by demonstrating that inference can be accelerated using consumer-grade memory, the company directly challenges the economic logic behind Nvidia’s data-center dominance and opens the door to massively cheaper, grid-scale model deployment.

Also tracking

Voice & Speech

What moved

Build more natural voice experiences with GPT-Live-1 in the API (OpenAI)

OpenAI’s GPT-Live-1 adds full-duplex, natural voice conversations, custom voices, and telephony support to the API, directly advancing real-time voice agents with more responsive and lifelike interactions.

Intelligent transcription with Gemini 3.5 Transcribe (deepmind.google)

DeepMind’s Gemini 3.5 Transcribe is pitched not as better speech-to-text but as smarter speech-to-text: the claim is transcription that understands what it’s processing rather than merely converting audio to words. That distinction matters because raw word accuracy has largely stopped being a differentiator in ASR — the competition has shifted to what the output can support downstream. Branding it under Gemini also fits Google’s pattern of routing useful capabilities through its flagship model family instead of shipping standalone utilities, which makes this a positioning play as much as a product one. In a crowded, fast-moving transcription market, owning the model layer beneath everyone else’s audio pipeline is the prize.