AI Daily · November 10 Tuesday

2026-11-09 08:00 — 2026-11-10 08:00 (UTC+8)

This issue includes 8 supplemental stories.

1. 01

   Nominal raises $240 million Series C at a $3 billion valuation

   Nominal announced a $240 million Series C at a $3 billion valuation on November 9, led by Lightspeed Venture Partners and co-led by Insight Partners. The company said it will turn four years of hardware test-data infrastructure into a Hardware Intelligence Platform for engineering teams. Bloomberg, reporting before Nominal’s official post, said the valuation is triple a funding round earlier this year.

   Read the full report

   Hardware-engineering software company Nominal said on November 9 that it had raised $240 million in a Series C at a $3 billion valuation. Its official blog said the round was led by Lightspeed Venture Partners and co-led by Insight Partners. The company’s official X account later posted a longer investor list that also included Lux Capital, Sequoia, General Catalyst, Founders Fund, and Lockheed Martin Ventures, among others.

   Nominal said that for four years it has mainly helped hardware teams work with test data and make decisions from it, and that this round will turn that foundation into a Hardware Intelligence Platform connecting physical systems, engineering data, model agents, and engineering applications. It also said its first purpose-built agent, Nominal Analyst, can draw on data and context in the platform to help with analysis. Bloomberg reported the same deal at 13:00 UTC (21:00 Beijing time), before the company’s official post, and said the valuation is triple what Nominal was worth in a funding round earlier this year. Bloomberg described the company as a startup building software to make manufacturing more efficient.

   The official announcement did not disclose revenue, customer numbers, or a breakdown of how the proceeds will be used. The date the round closed has not been made public.

2. 02

   Meituan releases LongCat-2.5 weights under an MIT license

   Meituan LongCat said on November 9 that LongCat-2.5 weights are now available under an MIT license on Hugging Face, alongside a W4A4 quantized version. Meituan describes the natively multimodal mixture-of-experts model as having 1.6 trillion total parameters and about 48 billion active per token, aimed at long-horizon agentic tasks. Open weights were only planned when the model launched in early November; benchmarks are Meituan's own.

   Read the full report

   Meituan LongCat posted on X at 22:05 Beijing time on November 9 that the weights for LongCat-2.5 were being released that day under an MIT license, linking to a Hugging Face repository; a W4A4 quantized version, LongCat-2.5-W4A4, was also published. The model card describes LongCat-2.5 as a natively multimodal mixture-of-experts model for long-horizon agentic tasks, with 1.6 trillion total parameters and about 48 billion activated per token.

   Meituan first announced LongCat-2.5 in early November, when it was available only as an online service and open weights were described as planned. The November 9 post quotes that earlier launch post and says the weights are now up.

   Benchmark results on the model card are Meituan's own tests, which the company labels as for reference only; no independent reproduction has been seen. The W4A4 version quantizes expert weights and activations to 4 bits for domestic Chinese accelerators; its throughput gains likewise come from Meituan's own load tests.

3. 03

   China's CAC launches unified filing system for generative AI and other new tech

   The Cyberspace Administration of China said on November 9 that its new Internet New Technology and New Application Information Service Filing System went live that day at beian.cac.gov.cn, handling filings for generative AI, algorithms, deep synthesis and blockchain in one place. The previous algorithm filing system and blockchain filing system were retired at the same time. The notice covers only the channel consolidation, not any change to filing requirements.

   Read the full report

   The Cyberspace Administration of China (CAC) issued a notice on November 9 saying it had built an Internet New Technology and New Application Information Service Filing System, which went live that day at beian.cac.gov.cn. The system handles filings for four categories: generative AI, algorithms, deep synthesis and blockchain.

   According to the notice, the former algorithm filing system and the blockchain information service filing system were retired at the same time. The CAC said the new system is meant to standardize filing work and make online submission and lookup of filing information easier.

   Providing generative AI services to the public in China requires completing the relevant filing; algorithm and blockchain filings were previously handled in separate systems. The notice does not mention any change to filing requirements or review standards, only the consolidation of the channel.

4. 04

   Adobe and Disney form multi-year AI alliance around Firefly Foundry

   Adobe and The Walt Disney Company announced a multi-year strategic alliance at Adobe MAX on November 9, naming Adobe Disney's official AI innovation partner. They will use Firefly Foundry to train custom generative models on Disney's own IP for production work across parks, studios and ESPN. Financial terms were not disclosed.

   Read the full report

   Adobe and The Walt Disney Company announced a multi-year strategic alliance at Adobe MAX on November 9 (US Eastern time), with Adobe becoming Disney's official AI innovation partner. According to Adobe's press release, Adobe will give Disney early access to products including Firefly, Firefly Foundry, Premiere and Photoshop, and dedicate engineering and research resources.

   The core of the deal is Firefly Foundry: the companies will train custom generative models on Disney's own IP, learning characters, settings and visual styles, for commercially usable production content. The release says Walt Disney Imagineering R&D is already using Firefly Foundry models, including one that turns hand-drawn sketches into 2D concept art. The alliance also covers parks and cruise development, Walt Disney Studios' StudioLAB and the ESPN Edge innovation center.

   Financial terms were not disclosed.

5. 05

   Google adds TPU 8tx and TPU 8i Max for confidential computing and long context

   Google announced on November 9 two additions to its eighth-generation TPU line: TPU 8tx for confidential inference, and TPU 8i Max, a dual-chip package for very long context windows. Google says this is the first time it has introduced four TPU systems in one year. Performance figures are Google's own without disclosed test conditions, and the blog does not give availability dates.

   Read the full report

   Google announced on its official blog on November 9 two new systems in its eighth-generation TPU line: TPU 8tx, for secure and confidential inference, and TPU 8i Max, for very long context windows and higher rack compute density. Together with TPU 8t and TPU 8i, launched in April, Google says this is the first time it has introduced four TPU systems in one year.

   TPU 8tx uses the same silicon as TPU 8t, paired with fifth-generation AMD EPYC host processors, keeping data and model weights encrypted and isolated in hardware during computation; it targets regulated industries and government customers. Google says latency is up to half that of TPU 8t with near-zero performance loss, but did not disclose test conditions. TPU 8i Max packages two TPU 8i chips together, doubling compute, on-chip SRAM and HBM bandwidth versus TPU 8i.

   Google also mentioned a TPU agent toolkit in preview and an open-source TPU Sync, and said TPUs now have nearly 100 customers. The blog does not state when the two new systems will be available.

6. 06

   ByteDance releases Seedream 5.5 with 100-round editing and transparent images

   ByteDance's Seed team released image model Seedream 5.5 on November 9. It says unedited regions stay unchanged across up to 100 rounds of editing and that the model natively generates RGBA images with transparency. It is rolling out on Jimeng, Doubao and Volcano Engine Ark. Claims of benchmark leadership are ByteDance's own; Sina Tech reports a 2K square edit costs about 0.28 yuan.

   Read the full report

   ByteDance's Seed team released its new image creation model Seedream 5.5 on November 9, saying it is rolling out on platforms including Jimeng, Doubao, Doubao Work and Volcano Engine Ark.

   The team lists three main upgrades: in multi-round editing, unedited regions stay pixel-level unchanged, allowing up to 100 consecutive edits on one image; native generation of RGBA images with an alpha channel for use as standalone assets; and more realistic lighting, materials and skin texture, with more accurate rendering of complex text. ByteDance says it leads mainstream models on text structure and text-to-image realism evaluations but did not publish evaluation details.

   According to Sina Tech, the Seedream 5.5 API charges for input and output images with no charge for text tokens; a 2K square image edit costs about 0.28 yuan, less than half of Seedream 5.0 Pro. ByteDance says a Pro version will follow.

7. 07

   Baidu unveils ERNIE Evolve automated research system for model optimization

   Baidu announced ERNIE Evolve on November 9, an automated research system that, given a goal, picks directions, edits code and runs experiments, keeping only changes that pass independent evaluation. Baidu says it raised ERNIE 5.5 Flash deterministic GPU inference throughput by 33.7% and single-request decoding speed by 14.9%. All results are self-reported and not independently reproduced.

   Read the full report

   Baidu announced the automated research system ERNIE Evolve on its official X account on November 9, calling it a pragmatic step toward recursive self-improvement. According to Baidu, given a clear research goal, the system chooses research directions, modifies code and runs experiments, and only improvements that pass independent evaluation and review are kept.

   Baidu says ERNIE Evolve reached state of the art on two public tasks, NanoChat AutoResearch and NanoGPT Speedrun. In ERNIE 5.5 Flash inference tests, it raised deterministic GPU inference throughput by 33.7% and GPU single-request decoding speed by 14.9%. QbitAI reported that in a low-latency scenario on Kunlunxin P800 chips, the system took about 168 hours to raise generation speed to 623.6% of baseline.

   All of these results are self-reported by Baidu and have not been independently reproduced. Baidu says it plans to extend the approach to training systems, model capabilities and evaluation.

8. 08

   Nikkei says Sakana AI is preparing an IPO; company says no near-term plans

   Nikkei reported on November 9 that Japanese AI company Sakana AI has begun preparing for an IPO, possibly as early as autumn 2027. Sakana AI responded on X that evening that it has no near-term IPO plans and that no formal decision has been made. The response did not directly deny the reported timeline or adviser details.

   Read the full report

   Japanese AI company Sakana AI posted on its official X account on the evening of November 9 that it has "no near-term IPO plans" and that its priority remains building a national AI champion for Japan; the Japanese-language part said no formal decision on an IPO has been made.

   The statement responded to media reports that day. Nikkei reported that Sakana AI had begun preparing for an IPO, possibly listing in Japan or the US as early as autumn 2027, and had hired Goldman Sachs as an adviser. Reuters then cited people familiar with the matter as saying listing preparations were under way; Sakana AI's PR department told Reuters no formal decision had been made.

   Sakana AI's statement denies a near-term listing and a formal decision, but does not directly deny the reported autumn 2027 timing or the adviser detail. Nikkei's full article is paywalled; its content here is based on its public summary and Reuters reporting.

9. 09

   India's Sarvam names former Google Cloud India head as president and CBO

   Indian AI company Sarvam announced on November 9 that Sashikumar Sreedharan, former managing director of Google Cloud India, has joined as president and chief business officer and joined its board. He will lead enterprise and government business plus marketing, finance and deployment engineering. Sarvam says more than 100 enterprises and government bodies build on its products.

   Read the full report

   Indian AI company Sarvam announced on November 9 that Sashikumar Sreedharan has joined as president and chief business officer and has joined its board. According to Sarvam's blog, he will lead the enterprise and government business, expand work with digital-native companies, developers and partners, and oversee marketing, finance, forward-deployed engineering, service delivery and business operations.

   Sarvam says Sreedharan was previously managing director of Google Cloud India and earlier led enterprise business at Microsoft India. Sarvam says more than 100 enterprises and government bodies are building on its full-stack products; that figure is the company's own.