🎉 Step 5 Preview has Day-0 support in vLLM. Sparse attention, MoE, multi-token prediction and vision, with serving work across Blackwell and Hopper. The work spans the sparse indexer, FP8 MegaMoE and cache transfer, with kernel results and further details at https://t.co/f9apB6vVGG
NVIDIA has added DLSS 4.5 controls to GeForce NOW, letting members adjust image quality and performance in supported games. The settings are available in the app on Windows, macOS, Chrome OS and Steam OS, with Ultimate members on GeForce RTX 5080-powered cloud rigs gaining options including Dynamic Multi Frame Generation up to 4X. Enshrouded 1.0 and Warhammer 40,000: Boltgun 2 are also streaming at launch, alongside 10 games joining the service this week.
Why it matters: The membership and platform details clarify how access to cloud graphics controls varies by tier, with the broadest options available on RTX 5080-powered Ultimate rigs.
Day-one support for Jais 3 is available in vLLM. Congratulations to @IFM_AI and collaborators on the launch. Arabic-native agentic AI, ready to serve with vLLM from day one. https://twitter.com/IFM_AI/status/2110641827778843010
Hikvision won the Industrial Automation Award at the 2026 China International Industry Fair on October 12 for its AI ultrasonic testing technology. The company says its ultrasonic foundation model analyzes raw signals to identify and measure defects, and can automatically interpret the full inspection process in typical scenarios. In one automotive parts project, defect identification accuracy was at least 99.5%, with a missed detection rate of no more than 0.5%.
Hikvision showcased more than 100 industrial digitization products and launched over a dozen new products at the 2026 China International Industry Fair, covering quality inspection, equipment monitoring, production operations, and workplace safety. Its production line SOP detection solution reportedly generates a new model from 1–8 images for deployment within minutes. A newly released semiconductor packaging 3D AXI inspection system combines X-ray imaging with an industrial inspection model, offering detail resolution below 1 micrometer and geometric magnification up to 2500 times.
Token Rhythm says more than 300,000 developers have joined its platform and has launched limited-time Singles’ Day pre-sale offers. A single recharge of at least 50 yuan on TokenRhythm API earns the same amount in model usage credits, with no cap on cumulative bonuses; the credits are permanently valid. The company also says previously gifted credits have been made permanently valid.
Modal crossed $1B in annualized run rate and raised a $750M Series D led by Accel at a $15.75B valuation, according to the company announcement cited by Latent Space. Usage grew more than 15x year over year. Modal’s Erik says Sandboxes alone reached a $300M run rate, with close to 3 billion sandboxes launched and support for a million concurrent sandboxes across uses including background agents and reinforcement learning.
Leiphone reports that some AI infrastructure specialists can command annual compensation of 7–8 million yuan as model companies seek lower computing costs. Interviewees identify inference optimization, model quantization, and kernel development as premium skills, while routine framework deployment attracts smaller rewards. StartLux co-founder and CTO Guo Quanwei estimates that only 5%–10% of AI infrastructure workers receive substantial premiums, challenging the impression that salary doubling is widespread.
Fish Audio Drama 3 is now supported in vLLM-Omni, offering voice cloning, speaker tags, expressive delivery through /v1/audio/speech, and streaming 44.1 kHz audio. The merged optimization PR reports 128 concurrent streams on one H200, 255 ms median time to first audio, and 222.2 seconds of audio generated per second; both stages share the GPU, averaged over 5 runs of 1,088 Seed-TTS-Eval English requests per run with pre-encoded voice references. Try current upstream: vllm serve fishaudio/drama-3 --omni --port 8091 Recipe: https://github.com/vllm-project/vllm-omni/blob/main/recipes/fishaudio/Drama-3.md
NVIDIA describes how Nordic countries are building a technology ecosystem around computing infrastructure, startups and industrial applications. AI cloud providers are projected to deploy 1 gigawatt of capacity across the region by the end of 2027, while Sweden’s Sferical One launched a shared supercomputer with 16 racks of NVIDIA DGX GB300 systems and 1,152 GPUs. The source also highlights Nordic companies using NVIDIA technology in recommendations, robotics, meeting assistance and other applications.
Artificial Analysis is adding AMD MI355X results to AA-AgentPerf, which replays real coding agent sessions with prompt caching and tool-call delays. On a single 8-GPU node, MI355X delivers ~15% to 40% more throughput per MW than B200 on DeepSeek V4 Pro 0813 between 35 and 75 tok/s per user. DeepSeek V4.1 Flash matches NVIDIA Blackwell performance for a single agent but trails as load rises; GLM-5.2/5.3 and Kimi K3 also trail the compared NVIDIA systems. Results use ROCm nightly and dev builds of vLLM and SGLang; Kimi K3 comparisons use NVIDIA multi-node configurations.
Aravind Srinivas shared a write-up on how Perplexity’s sandbox platform handles millions of executions a day while using multiple machines to lower serving costs. https://t.co/lmRrwROweu In the quoted post, @perplexity_ai says SPACE has an error rate below 0.0002% and uses cryptographic identity to keep requests tied to the right execution as sandboxes pause, resume, and move. https://t.co/jH6TX2cgI9
AWS explains how to build a document processing system using Amazon Bedrock AgentCore and eight specialized agents that classify documents, extract data, validate results, and save refined extraction instructions for reuse. In its purchase order example, processing takes 3 minutes 49 seconds for an unfamiliar format and 1 minute 30 seconds for a subsequent document using stored instructions. The guide includes deployment steps and explains how to customize schemas, validation logic, and prompts.
Our sandbox, SPACE, handles millions of executions a day with an error rate below 0.0002%. We’re sharing how it uses cryptographic identity to keep requests tied to the right execution as sandboxes pause, resume, and move. Read more: https://t.co/jH6TX2cgI9
Aravind Srinivas describes a user's project in which Perplexity Computer ported a Windows-only mechanical keyboard app with no Linux support or documentation to a native Rust app for Linux. Computer downloaded the binary in its cloud sandbox, decompiled it with headless Ghidra, and read the code to understand how it drives the keyboard. It rebuilt the app piece by piece with music-synced lighting over PipeWire, a 3D view, key remapping, macros, and profiles. https://www.reddit.com/r/GeminiAI/comments/1x4vvoi/my_mechanical_keyboards_software_only_runs_on/ https://github.com/trenbolone1122/aula-f75-linux
Day-0 support for Ling-3.1-flash is here in vLLM! ~560B total parameters, ~25B active per token, and up to 1M context. From a rough brief to websites, mobile UIs, and interactive prototypes, serve it on your own infrastructure with vLLM. Congrats to @AntLingAGI! https://twitter.com/AntLingAGI/status/2110390124743970952
Epic Games announced a preview of its redesigned Epic Games Launcher for Windows, Mac, and Linux beginning today, describing it as faster and more responsive. Epic executive Martin Keely told The Verge that agents wrote and reviewed all the code under the development team’s supervision, with AI helping replicate existing functionality and support Linux. The old launcher remains available, and Epic has no specific timeline for requiring migration.
Amazon SageMaker JumpStart provides a workflow to deploy a pretrained model, run a repeatable performance test with the SageMaker Python SDK, and delete the endpoint afterward. The test measures latency, estimated output length, and single-request generation rate using sample prompts. Model Hub discovery and preconfigured endpoint options simplify browsing and deployment, while a deep link integration lets users deploy from Hugging Face model discovery in a single click.
AWS’s post explains how to build an AI agent system with Amazon Bedrock AgentCore to create client PowerPoint presentations from enterprise documents. In internal testing, the workflow completed in under 90 seconds; it uses a supervisor agent, RAG, and specialized tools for presentation generation and summarization. The post includes a public repository with code assets.
Ant Group’s embodied AI company Ant Lingbo has signed strategic cooperation agreements with Longsheng Weirui, Dema Technology and De’er Xiangwang to develop robotics solutions for industrial and logistics settings. The work includes industrial loading and unloading, automotive factory digitalization, and flexible parcel sorting. At the China International Industry Fair, robots from Tihoo and Stardust Intelligence equipped with LingBot-VLA 2.0 demonstrated industrial loading and unloading and flexible sorting.
STEPX Neo has launched at 3999 yuan with the agent-native Step AOS, personal agent Amoo, and a STEPX Token Plan providing 10 billion tokens monthly for 12 consecutive months. After two weeks with a sample device, Jay describes syncing vocabulary lessons to a Feishu calendar through CLI tools and building an AI news app in Step Studio. The article also outlines tool-call logs, confirmation requirements for high-risk actions, and recovery options for agent-deleted data.
Tencent released integrated versions of WeCom and WorkBuddy on October 14, connecting workplace context through a shared WeCom account system. WeCom 5.0.12 adds WorkBuddy as a conversational bot in enterprise contacts, while WorkBuddy 5.8.0 supports WeCom login and authorized access to messages, documents, meetings, and cloud storage. Administrators can manage accounts, usage, AI compute purchases, and access permissions; integration is available to verified WeCom enterprises with limited free trial slots.
SenseTime’s AI platform, based on the upgraded “Cloud Universe Starry Sky” model, developed and delivered three agents for Shanghai’s first batch of government agents under its “Quantum City” initiative. They handle flood-control dispatch, spatial analysis, and planning updates. In the Suzhou Creek basin, the “Yuzhi Heluo” agent improved dispatch efficiency by 80%, alongside reported gains of 65% in spatial computing and 70% in dynamic simulation.
@lightseekorg reports 685 output tokens/s per user for Qwen3.8-Flash-Next. Congrats to @lightseekorg on the impressive TokenSpeed results! 🚀 Great to see Qwen3.8-Flash-Next reaching 685 output tokens/s per user on Nvida GPUs. Excited to see the community unlock more of Qwen’s inference potential!
Tencent WorkBuddy Enterprise added an observability center to its admin console on October 13 and opened it for public beta. It provides real-time monitoring, session tracing, and execution trace details covering usage, credit consumption, response times, and errors. Administrators can break down consumption by department and model, investigate specific execution nodes, and adjust Agent prompts, models, and tools at https://www.workbuddy.cn/admin/overview.
Drawing on Anthropic’s engineering blog, Baoyu explains why the company replaced git with Scion as daily merged code changes grew from around 2,500 in May to more than 30,000 in September. Scion retrieves files on demand, moves operations to the server, and uses layered caches with hit rates above 99%. Anthropic reports that median agent workspace startup fell from 7.1 minutes to 5.1 seconds, while emphasizing that Scion serves exceptionally large monorepos and is not a general-purpose git replacement.
Why it matters: The comparison illustrates how assuming continuous server access changes source control architecture, connecting on-demand file retrieval and server-side operations to the infrastructure demands of agent-generated code.
Profiling the unmodified web-edit app with Python 3.15’s built-in `profiling.sampling` found that preserving mixed line endings dominated the editing workload. `preserve_endings` appeared in 64.8% of 15,296 CPU stacks, while `SequenceMatcher.find_longest_match` contributed 38.3% of leaf samples; 95.1% of stacks came from AnyIO’s worker thread, making all-thread sampling essential.
Simon Willison highlights Anthropic’s explanation of Scion, its custom alternative to Git, as merged code changes grew from around 2,500 per day in May 2026 to over 30,000 per day in September. Anthropic reports almost 100 GB of main-branch history and roughly 25 million daily CI checkouts; seven people built the system, completing migration in five months. Willison speculates that Fable/Mythos class models contributed to the increased coding velocity and connects Anthropic’s experience with adversarial review to his own, including the need for engineers to steer models away from over-engineering.
Why it matters: Anthropic’s checkout volume and repository size offer concrete context for understanding how increased coding activity can put pressure on the infrastructure supporting code review and CI.
@leadinginds says Ramp spend data shows Perplexity, Wispr Flow, and Claude Code each bought by 87%+ of buyers in their categories. Inference, GPU capacity, and AI-first GTM remain fragmented. 👀 https://twitter.com/leadinginds/status/2110080629773635595
The guide explains how Anchor Browser integrates with Amazon Bedrock AgentCore payments to handle x402-protected pages and APIs within the same browser session. Setting session.x402 to true enables automatic payments subject to configured spending and expiration limits, while wallet credentials remain outside the agent runtime. Existing browser automation code can continue unchanged; rejected payments or unsupported challenges leave the original HTTP 402 response intact.
Why it matters: The integration illustrates a transferable architecture for paid agent workflows by placing payment signing and budget enforcement outside the agent runtime while preserving browser session context.
Artificial Analysis now publishes cache hit rates for tracked model endpoints and provider-level Cost per Task for open weights models. Cache hit rates are measured using multi-turn agentic coding tasks; provider cost comparisons use the same token counts from each model's Intelligence Index run. GLM-5.3 (max) costs vary ~9x, from $0.92 on Makora to $8.68 on Nebius, largely because of input pricing despite both having high cache hit rates. Across 21 providers (25 endpoints), cache hit rates range from 41% to 100%.
Flox Intelligence makes on-site AI devices that identify nearby wildlife and emit species-appropriate sounds to deter it. The company says about 150 devices are operating across North America, Europe, and Australia, and it expects about $1.5 million in revenue by the end of 2026. Flox is opening a $10 million Series A to expand deployments and extend the system beyond its current wildlife use cases.
Origin, Cursor’s code hosting platform, now supports over 20 integrations. Connect your repos to tools for CI, deployments, code review, security, monitoring, and project management. https://t.co/Xq8LpS6Fa7
GitHub explains how its secret scanning pipeline combines regex candidate detection with a fine-tuned ModernBERT classifier to distinguish real credentials from lookalikes using surrounding context. Compared with the previous LLM pipeline, offline evaluation showed 13% higher precision and 19% higher recall; a customer A/B experiment showed 11.5% higher repository owner-weighted acceptance. Median post-push scan execution time fell 86%, and serving costs dropped more than 98%, making it practical to extend generic secret detection to push protection.
Why it matters: GitHub’s pipeline illustrates a transferable design pattern for narrow classification tasks, combining deterministic candidate generation with a compact contextual model trained on realistic hard negatives.
AWS explains two ways to activate managed entitlements for Amazon Bedrock private pricing offers across AWS accounts. A lightweight script supports one-time activation of trusted received licenses, while AWS CDK automation processes agreement events, verifies sellers against an allow-list, and creates and activates distributed grants. Negotiated pricing applies only after grants reach ACTIVE or WORKFLOW_COMPLETED; the guide also covers checking grant status and validating charges in AWS Cost Explorer after billing data appears.
Customizing Vertical Voice Agents with ASR and LLM Fine-Tuning https://t.co/dBSfkUiKeX
Orca Security partnered with AWS to build Discovery AI on Amazon Bedrock, translating natural language questions into validated Sonar queries across 900+ cloud infrastructure and application models. The architecture combines example retrieval and reranking, relevant schema selection, agentic generation, and validation, with domain experts continuously updating the example bank. Compared with its initial baseline, Orca reports 90 percent lower costs, 93 percent fewer input tokens, and 75 percent lower latency, with response times falling from approximately 60 seconds to 15 seconds.
Why it matters: The architecture offers a transferable approach to querying large schemas by retrieving validated examples, selecting relevant schema fragments, and feeding validation errors back into generation.
Anthropic’s Engineering blog shares how it built a new version control system and fully migrated its monorepo from git in five months. In May, it merged around 2,500 code changes per day; in September, it merged more than 30,000 per day. That growth created bottlenecks in developer infrastructure built for human rather than agent scale: https://www.anthropic.com/engineering/redesigning-source-control
The same story is shown in the picks as 《Why Anthropic built Scion to replace git for its monorepo》
Satlyt has raised an $8 million seed round led by Non Sibi Ventures to build software for running AI models across different companies’ satellites. An earlier deployment of Google DeepMind’s Gemma reduced a transmission about onboard software errors by more than 60%. Its software is set to launch Thursday on a SpaceX rocket alongside Google’s Project Suncatcher prototype, and the company expects to test shared computing across two satellites next year.
AWS explains three fused MoE optimizations used in Amazon SageMaker AI, reporting up to 10 percent higher throughput and up to 9 percent faster Time to First Token (TTFT). On H200, swapping matrix operands improved throughput by 2–4 percent, while routing FP8 tensor-quantized models to Triton improved throughput by 10 percent. Eliminating unnecessary Triton JIT recompilation improved TTFT by 9 percent at 8-way tensor parallelism; the post identifies availability in vLLM and the AWS LMI v28 container.
Why it matters: The optimization examples offer a transferable approach to MoE serving performance by examining matrix layout, backend selection, and compiler specialization as separate sources of throughput and latency overhead.
