AI delivery meets the memory wall and the real world.

Beam moves from announcement to open weights

16 Oct · Recent development

Reflection released Beam weights under Apache 2.0, with BF16, FP8 and NVFP4 formats plus inference, fine-tuning and agent integrations. The coding and reasoning model has about 501 billion total parameters, 23 billion active per token and a one-million-token context, according to its model card.

The useful comparison is integration effort, deployment requirements and cost per completed task. A small active parameter count does not mean the full weights fit on an ordinary laptop; vendor benchmarks are not a substitute for workload testing.

Reflection: Beam is now available

Beam model card

Anthropic proposes a 5% equity commitment to a public-benefit fund

19 Oct Beijing · New in this edition; website dated 18 Oct

Anthropic announced a wholly owned LLC for public-benefit work and a proposed 5% equity commitment, subject to stockholder approval. The board may add two percentage points for each 50% increase in company value above a specified baseline. Labor-market impacts are one potential focus.

The next concrete measures are governance and actual distributions. This is an equity arrangement, not cash already paid out; the detailed valuation milestones and governance structure are still to come.

Anthropic: Investing in the public benefit

Claude Code · Anthropic

2.1.303 · 2026-10-19 · stable

Released at 06:26 Beijing. Notes list bug fixes and reliability improvements, without a new capability announcement.

Claude Code 2.1.303 release

Checked 2026-10-19T03:59:02.801377+00:00

DeepSeek · DeepSeek

V4.1-Flash-1017 · 2026-10-17 · stable

Updated on web, app and API; deepseek-flash and pricing stay unchanged. Search/Fetch API opened on the same date.

DeepSeek change log

Checked 2026-10-19T03:59:02.182909+00:00

Reflection Beam · Reflection

Beam · 2026-10-16 · stable

Apache 2.0; BF16, FP8 and NVFP4 weights, with a one-million-token context. Evaluate deployment and performance on your own tasks.

Reflection: Beam is now available

Beam model card

Checked 2026-10-19T03:41:43.503672+00:00

Cursor · Cursor

Cloud agent canvases · 2026-10-16 · stable

Cloud agents can deliver shareable interactive canvases. The separate Environments API is in beta.

Cursor: Canvases from cloud agents

Checked 2026-10-19T03:59:04.061278+00:00

Mistral Timestral · Mistral

5M / 28M / 123M / 209M · 2026-10-15 · stable

Multivariate probabilistic forecasting, including CPU inference. Code uses Apache 2.0; the Modified MIT weights licence excludes companies above $20 million in prior-month revenue without separate permission.

Mistral: Introducing Timestral

Timestral weights licence

Checked 2026-10-19T03:59:04.155423+00:00

Physical Intelligence · Physical Intelligence

Pi API beta · 2026-10-13 · beta

Upload robot demonstrations, fine-tune on managed infrastructure and use hosted inference. Access remains a beta application.

Physical Intelligence: Pi API beta

Checked 2026-10-19T03:58:58.988680+00:00

SK hynix connects memory pools, SSDs and the KV cache

Published 16 Oct · OCP demonstrations on 12–15 Oct

At OCP, SK hynix demonstrated CXL pooled memory and its MARS memory-centric rack, including KV-cache sharing and DRAM/SSD reuse. It described proof-of-concept work with NAVER Cloud and system demonstrations with partners including AMD. The portfolio also spans HBM, server DRAM and enterprise SSDs.

This broadens the storage question beyond HBM: long-context agents need capacity and reusable data as well as bandwidth. Demonstrations and proofs of concept do not establish deployment at scale or universal speed gains.

SK hynix: Memory-centric systems at OCP

Digit 5 targets nine-minute charging and longer working days

15 Oct · Seven-day watch

Agility selected Nyobolt for Digit 5 batteries. Its announced design targets 90 minutes of runtime and nine minutes of charging, moving the run-to-charge ratio from Digit 4’s 2:1 to 10:1 and targeting more than 20 working hours per day. Early access is expected in the first half of 2027.

These figures matter to shift planning and task economics, but remain company targets. Real-site task success, human intervention and battery life are the relevant acceptance tests.

Agility: Digit 5 fast-charging partnership

Pi API beta makes the robot-learning loop accessible to developers

13 Oct · Seven-day watch

Physical Intelligence opened applications for its Pi API beta. Developers upload demonstrations, fine-tune models on managed infrastructure and run inference through a hosted gateway. Its examples include heavy machinery, kitchens and manufacturing; it also acknowledges connectivity issues in some applications.

The shift is a practical path for robot companies to iterate their own policies. This is a beta workflow for physical robots, and the company’s examples do not establish performance across all hardware and sites.

Physical Intelligence: Pi API beta

NVIDIA opens rack designs as power and cooling become shared specifications

12 Oct · Seven-day watch

At OCP, NVIDIA contributed Vera Rubin rack and compute-tray designs. MGX 1.2 covers blind-mate liquid cooling, rack interconnects and designs supporting up to 250 kW per rack. It also introduced open software for cluster readiness and rack management; partners outlined 800 V DC designs.

Standard interfaces and operations could reduce integration work. The 250 kW figure is a rack-design capability and 800 V is a power architecture, not newly commissioned data-centre capacity.

NVIDIA: Open rack and facility specifications

AWS Transform adds Azure-to-AWS code conversion

16 Oct · Recent development

AWS made Azure-to-AWS SDK conversion for .NET applications generally available in Transform, covering Functions, Blob Storage and Queue Storage mappings. Cosmos DB-to-DynamoDB conversion remains in preview. The service is available in eight listed AWS Regions.

Teams can review and adjust the plan before conversion. The value is less repetitive rewriting; generated code alone does not complete a production migration or verify data consistency, permissions and business behavior.

AWS: Azure-to-AWS code conversion

China’s export proposal distinguishes API services from technology transfer

16 Oct · Recent development

MOFCOM and the science ministry proposed adding 23 key AI technologies to the technology-export catalogue. The official Q&A distinguishes transfers of restricted technology for overseas deployment from API, app/web and SaaS service access. Temporary pre-export registration for covered technologies applies from the announcement.

Cross-border teams need to check the covered technology and transfer method. The catalogue revision is a proposal, not a blanket restriction on AI services; temporary registration and final licensing rules are distinct.

MOFCOM: Technology export catalogue Q&A

The Dublin Declaration links AI sovereignty, compute and adoption

Adopted 14 Oct; published 15 Oct · Seven-day watch

Participating EU member states set out seven cooperation priorities, including AI safety, public-sector and industrial adoption, sovereignty, infrastructure and compute, sustainability and international cooperation. Ireland published the declaration on 15 October, recording adoption on 14 October.

This is a cooperation declaration, not a newly effective EU regulation. Procurement, compute investment and implementation are the next signals to track.

Irish government: Dublin AI Declaration

CXMT forecasts a nine-month profit as the memory cycle strengthens

17 Oct · Recent development

CXMT forecasts nine-month revenue of RMB265–270 billion and net profit attributable to shareholders of RMB131–132.5 billion, versus a RMB5.28 billion loss a year earlier. It attributes the change to DRAM demand and prices, utilization and improved output per wafer.

These are unaudited preliminary estimates released on 17 October, not the final quarterly report. The announcement postdates the 16 October close used here, so it cannot explain that earlier day’s gain.

CXMT: Preliminary nine-month earnings forecast

TSMC reports $46.72 billion revenue and a higher Q4 range

15 Oct · Seven-day watch

TSMC reported Q3 revenue of $46.72 billion, above its $44.6–45.8 billion guidance, with a 66.4% gross margin. Its Q4 revenue guidance is $52.3–54.3 billion and gross-margin guidance is 65–67%.

The release separates realized demand from the next quarter’s expectations. Total revenue includes multiple end markets and should not be presented entirely as AI revenue.

TSMC: Q3 2026 results

RapidCanvas raises $29 million for a platform-and-engineers model

16 Oct · Recent development

RapidCanvas announced a $29 million Series B led by Vertex Ventures SEA & India, with Hitachi Ventures and other investors participating. Total funding reaches $52.5 million. Its approach pairs an AI platform with field engineers to build applications and manage them in production.

The financing backs delivery and ongoing operations. The company also reports eightfold annual recurring revenue growth over six quarters; that ratio does not supply a comparable revenue amount or profit figure.

RapidCanvas: $29 million Series B