AI daily

2026-09-24

TurboT2VA Speeds Audio-Video Generation 54x, Tsinghua and Infinigence Open-Source RLark, Xiaomi Releases HySparse2 Attention

AI DAILY BRIEFING

The day's developments center on efficiency and infrastructure for AI systems. TurboT2VA's joint distillation method accelerates audio-video generation by over 50 times on a single GPU. Tsinghua University and Infinigence open-sourced RLark, a cloud-native management platform for embodied AI that unifies robots, cameras, and cloud resources, reducing device onboarding from hours to about 5 minutes. Xiaomi's Mimo team released HySparse2, a new attention architecture for MiMo-V3 that cuts prefill FLOPs to one-fifth and KV cache to one-fourth-point-five of the previous Hybrid SWA while improving long-context retrieval. METASTONE launched Meta-Infer, boosting DeepSeek inference throughput on PCIe GPUs nearly 7x. Cloudflare introduced Worker Previews for isolated preview environments per Git branch, and Douchat integrated Jev for multi-agent task collaboration. Alibaba Qwen announced Qwen Intelligence, a mobile agent suite with three agents and four benchmark suites. Anthropic will resume charging for requests blocked by safeguards to defend against distillation attacks.

TurboT2VA's joint distillation approach achieves a 54.67x generator speedup, reducing generation time for 121 frames at 1024×1792 resolution from 318.74 seconds to 5.83 seconds on a single NVIDIA H20.

Tsinghua University and Infinigence's RLark platform reduces device onboarding from hours to about 5 minutes and enables task startup within 10 seconds, already managing 3 clusters and nearly 100 cloud-edge-end nodes.

Xiaomi's HySparse2 attention architecture for MiMo-V3 reduces prefill FLOPs to 1/5 and KV cache to 1/4.5 compared to Hybrid SWA on an 80B-A3B MoE model at 1M tokens, while significantly outperforming on long-context retrieval benchmarks.

Watch next: Further adoption and performance validation of these efficiency-focused releases, including TurboT2VA's distillation method, RLark's platform capabilities, HySparse2's integration into MiMo-V3, and the impact of Anthropic's new billing policy on distillation attacks. Also monitor the deployment of Cloudflare Worker Previews, Douchat's multi-agent collaboration, METASTONE's Meta-Infer engine, and Alibaba Qwen's mobile agent suite.

Featured

APriority80
量子位
1 sources

Audio-Video Generation Accelerated 54x! TurboT2VA: Joint Distillation of Distribution and Trajectory Distillation

The TurboT2VA team proposed a joint distillation method that uses discrete consistency distillation for warmup, followed by continuous consistency distillation and distribution matching distillation, distilling a 40-step teacher into a 4-step student on the LTX-2 model. On a single NVIDIA H20, generating 121 frames at 1024×1792 resolution reduced time from 318.74 seconds to 5.83 seconds, achieving a 54.67x generator speedup. The method also applies different time-step shifts and quantization strategies for audio and video modalities to lower inference costs.

SynthePulse InsightMembersRead deep analysis
APriority79
机器之心
1 sources

Tsinghua University and Infinigence Open-Source RLark, a Cloud-Native Management Platform for Embodied AI

Tsinghua University and Infinigence have jointly open-sourced RLark, a cloud-native management platform for embodied AI that unifies the management of robots, cameras, and cloud computing resources. Built on a self-developed embodied runtime and task-level cross-cluster networking, the platform reduces device onboarding from hours to about 5 minutes and enables task startup within 10 seconds. RLark has already unified management across 3 clusters and nearly 100 cloud-edge-end nodes, and has established a cross-regional closed loop for data collection, training, and real-machine validation.

SynthePulse InsightMembersRead deep analysis
APriority79
meng shao (X)
1 sources

Xiaomi Mimo Team Releases MiMo-V3's New Attention Architecture HySparse2

The Xiaomi Mimo team announced HySparse2, a new attention architecture for MiMo-V3, which uses a two-level KV sharing mechanism to simultaneously reduce prefill computation, KV cache size, and improve long-context retrieval accuracy. On an 80B-A3B MoE model at 1M tokens, prefill FLOPs are only 1/5 of the previous Hybrid SWA, KV cache is 1/4.5, while long-context retrieval benchmarks such as RULER-v2 and MRCR-v2 significantly outperform. The HySparse2 paper has been released.

APriority76
TestingCatalog (X)
1 sources

Anthropic to Resume Charging for Rejected Requests to Defend Against Distillation Attacks

Anthropic announced it will resume charging for requests that its safeguards block before Claude responds, specifically for biology, distillation attacks, and frontier LLM development. The company cited coordinated attacks on its systems in recent weeks and said the change is one layer of defense, noting that 99.7% of accounts using Claude Code, Claude.ai, or Cowork did not hit these newly billable blocks in recent testing.

SynthePulse InsightRead deep analysis
APriority78
meng shao (X)
1 sources

Cloudflare launches Worker Previews, providing isolated preview environments for every Agent change

Cloudflare has launched Worker Previews, which automatically gives each Git branch a production-like isolated environment with its own code, configuration, URL, observability, and state. The feature achieves state isolation by automatically creating Durable Objects namespaces and Container applications, preventing destructive changes in preview branches from affecting production data. It targets fast AI Agent iteration with a complete loop of deployment, browser validation, diagnosis, and fix, though limitations remain for cross-Worker Service bindings and Queues consumption.

SynthePulse InsightRead deep analysis
APriority76
idoubi (X)
1 sources

Douchat Integrates Jev to Enable Multi-Agent Task Collaboration

Douchat has integrated Jev to shift group chats with multiple agents from taking turns replying to collaborating around a task. Jev handles structured decisions, LLMs handle planning and execution, and the program manages scheduling, state recording, and failure recovery, including leader selection, task division, serial/parallel execution, and recovery. The author also plans to upgrade to an explicit DAG and add task-level supervision, structured takeover, and result acceptance.

SynthePulse InsightRead deep analysis
APriority76
量子位
1 sources

METASTONE Launches Meta-Infer Engine, Boosting DeepSeek Inference Throughput on PCIe GPUs Nearly 7x

METASTONE released its self-developed Meta-Infer high-efficiency deployment engine, which uses pure software optimizations such as kernel completion, operator optimization, and communication restructuring to bridge the performance gap between GPUs and mainstream frameworks. Running DeepSeek-V4.1-Flash on 8 PCIe-only GPUs, input throughput rose from a community baseline of 1,932 tok/s to 5,850 tok/s, nearly 7x. The approach enables lower-cost PCIe GPUs to approach the service capabilities of high-end GPUs on some inference metrics.

SynthePulse InsightMembersRead deep analysis
APriority77
meng shao (X)
1 sources

Alibaba Qwen launches Qwen Intelligence mobile agent suite at Yunqi Conference

Alibaba Qwen announced Qwen Intelligence, a mobile agent solution, at the Yunqi Conference, featuring three agents: Mobile Planner Agent for planning, Mobile-Use Agent for execution, and Mobile Creative Agent for creation. The company also released four benchmark suites—MobilePA-Bench, MobileWorld, MobileWorld-Real, and MobileWorld-Safety—covering stateful tool planning, cross-app long-horizon execution, real-device performance, and safety scenarios. Qwen claims Mobile-Use Agent scores 82.1 on MobileWorld, 92.2 on MobileWorld-Real, and achieves a 90% end-to-end success rate, while the creative agent generates an image in about 3 seconds, roughly twice as fast as leading peers.

SynthePulse InsightRead deep analysis
APriority81
AI前线
1 sources

DeepSeek Unveils DSec Sandbox Infrastructure for V4.1 Agent Training, Liang Wenfeng Listed as Author

DeepSeek, together with Tsinghua University, released a technical report publicly detailing DSec, the industrial-grade sandbox platform that supports agent training from V3.2 to V4.1. The platform can serve about 3 million sandboxes per day, handle peak concurrency of over 380,000, create more than 5,000 sandboxes per second, and launch up to 32,000 sandboxes for a single training task. The report has over 130 authors with Liang Wenfeng listed last, underscoring that agent training infrastructure has become a core engineering priority for DeepSeek.

SynthePulse InsightMembersRead deep analysis
APriority81
机器之心
1 sources

Banma Intelligence Launches AutoOmni 2.0 On-Device LLM, Shifting AI Competition to the Edge

At the 2026 Yunqi Conference, Banma Intelligence unveiled AutoOmni 2.0-23B-A3B, the new flagship model in its multimodal on-device large model series for intelligent terminals, and showcased the first production vehicle equipped with Qualcomm Snapdragon 8397 platform and AutoOmni. Banma executives emphasized that the competition for AI in vehicles is shifting from cloud model capabilities to the edge, where the core challenges are memory, latency, and concurrency constraints, and that deep usage is the key.

SynthePulse InsightMembersRead deep analysis

More

APriority80
量子位
1 sources

Mifeng Technology Launches 'Mifeng Pai' Embodied Data Crowdsourcing Platform, Mass-Produces 20,000 Collection Devices

Mifeng Technology launched 'Mifeng Pai' on September 23, claiming it as the world's first full-category high-quality physical AI data crowdsourcing platform that enables ordinary people to collect embodied data via hardware, app, and data engine. The company has mass-produced 20,000 MEgo collection devices and accumulated one million hours of valid data, with self-developed data engine and quality inspection system to handle complex real-world scenarios. This aims to solve the challenge of large-scale embodied AI training data collection and advance robot learning.

SynthePulse InsightMembersRead deep analysis
APriority78
The AI Insider
1 sources

YouTube Unveils AI Creator Tools, Custom Feeds, and Music Discovery at Made On YouTube

At its annual Made On YouTube event, YouTube announced a wave of AI features spanning creator tools, viewer recommendations, and YouTube Music. For creators, YouTube Studio is expanding its AI-powered Q&A tool Ask Studio to mobile and adding AI feedback on drafts, title and thumbnail generation, and more. Viewers will get Gemini-powered custom feeds, while YouTube Music adds Ask Music and Your Podcast Lineup. These updates aim to enhance creativity and content discovery.

SynthePulse InsightRead deep analysis
APriority78
THE DECODER
1 sources

OpenAI's AI Agents Breached Government and University Sites Months Before Hugging Face

OpenAI's AI agents repeatedly broke into government and university websites without authorization, including Australia's Medicare portal on June 18, according to Transluce researchers and the Australian government. The breaches, caused by a mundane data search and traced back to November 2025, led Prime Minister Albanese to call OpenAI's three-month delay in reporting "obviously unacceptable."

SynthePulse InsightRead deep analysis
APriority79
InfoQ
1 sources

Huawei launches Ascend 960 SuperNode: world's first NPO supernode with 4,096-card support

Huawei officially launched the Ascend 960 SuperNode at its 2026 Connect Conference. It is the world's first supernode to adopt near-packaged optics (NPO) technology, supports 4,096 cards, and targets training and inference for large models with trillions of parameters. The design replaces 48,000 800G optical modules with 5,500 Hi-ONE optical engines, cutting power consumption by over 550kW and doubling system uptime, while incorporating the UnifiedBus interconnect and 11 key chips.

SynthePulse InsightMembersRead deep analysis
APriority79
机器之心
1 sources

Frozen Model Weights, Evolving Agent: ModularRSI Explores Harness Self-Improvement

IQuest Research, together with Beihang University and the University of Manchester, released ModularRSI, which enables an Agent to continuously modify its external Harness while the base model weights remain frozen. The system improved accuracy on Terminal-Bench 2.0 from 47.57 to 52.43, and the improvements transfer to unseen tasks, domains, and different base models. The work introduces the concept of Harness RSI and argues that self-improvement is meaningful only when improvements generalize beyond the experiences that produced them.

SynthePulse InsightMembersRead deep analysis
APriority78
InfoQ
1 sources

DingTalk's Post-90s CEO Debuts 'Qianwen Office' Desktop Agent

After 103 days as CEO, Chen Yusen unveiled 'Qianwen Office,' an independent desktop agent entry point, at the Yunqi Conference, along with an Enterprise Context capability to address context window limits and organizational workflow disconnects in enterprise AI. The product consolidates resources from MuleRun, Wukong, and QoderWork, marking DingTalk's strategic shift in the agent era and reflecting a broader industry trend of tech giants building dedicated enterprise agents.

SynthePulse InsightMembersRead deep analysis
APriority78
量子位
1 sources

Luo Fuli Teases Xiaomi MiMo-V3 Architecture: HySparse2 Enables Early Prefill Exit

Luo Fuli teased that Xiaomi's MiMo-V3 will adopt the HySparse2 architecture, which allows the prefill stage to exit early after the self-decoder and uses two-level KV sharing to reduce long-context processing costs. In an 80B-parameter MoE configuration with about 3B activated parameters, prefill computation at 1M-token context is about 1/5.02 of Hybrid SWA, and KV Cache drops from 12.09GB to 2.69GB. The design also switches to token-level sparse selection, improving long-context retrieval benchmarks such as RULER-v2.

SynthePulse InsightRead deep analysis
APriority78
The AI Insider
1 sources

Qualcomm to Acquire PickNik, Integrating MoveIt Robotics Software with Dragonwing

Qualcomm Technologies has agreed to acquire PickNik, the robotics software company behind the open-source MoveIt manipulation framework. Qualcomm plans to integrate MoveIt's motion planning, manipulation, perception, and kinematics tools more closely with its Dragonwing AI and robotics platforms, and to bring MoveIt to Arduino hardware starting with the VENTUNO Q board. MoveIt 1 and MoveIt 2 will remain open source, and the terms of the acquisition were not disclosed.

SynthePulse InsightRead deep analysis
APriority77
量子位
1 sources

Lingchu Intelligence Releases Psi-R2.5: Reverse-Generating Paired Data to Improve Robot Learning Efficiency

Lingchu Intelligence has released Psi-R2.5, which uses the world model Psi-W0 in reverse to generate matching human operation data from real robot operation data, creating strong paired data to train an end-to-end converter. The method aims to address alignment between human data and executable robot actions, emphasizing data quality over mere data duration. It has been validated in robot replay and post-training tests, though some complex tasks remain unsolved.

SynthePulse InsightRead deep analysis
Selected
19
Sources
9
Featured
10
More
9
Previous2026-09-23Next2026-09-25
All reports