TaichuAI's 10B Multimodal ZDTaichu5.0-9B — Spatial Reasoning and Agents, Open Weights Included
China's TaichuAI has released ZDTaichu5.0-9B, a multimodal model that puts spatial reasoning, embodied AI, and agent use front and center alongside general vision understanding. It pairs a Qwen3.5-9B language backbone with an NVIDIA C-RADIO-family vision encoder (C-RADIOv4-H) and supports contexts of up to 128K tokens. It accepts text, multiple images, and video at arbitrary resolutions, and it showed up on Hugging Face's trending list shortly after release.
The distinguishing design choice is that it layers spatial, embodied, and agent capabilities on top of general vision understanding (images, documents, charts, OCR, visual math) rather than trading it away. The training scope spans everything from basic spatial relations like left/right and front/back, through multi-view association, 3D scene understanding, and mental rotation, all the way to affordance understanding and action planning aimed at VLA. In the model card's comparisons it ranks first in spatial ability among open VLMs in the same 10B class (Qwen3.5-9B, STEP3-VL-10B, gemma4-8B-E4B), posting numbers such as ViewSpatial 62.50, MindCube-tiny 78.27, and SparBench 51.82.
The agent-side numbers stand out as well. On TAU2-Bench it scores 87.7, ahead of the closed models in the comparison table (GPT-5.2 at 87.1 and Gemini 3 Pro at 85.4), while Claw-Eval 71.4 and IFEval 93.7 are also the best among the compared 10B-class models. On the embodied side, however, it trails on ERQA with 48.00 against Gemini 3 Pro's 66.00, so the claim is strictly “first among 10B-class open models.” For reasoning, the model also introduces a mechanism called Entropy-Gated Adaptive Recurrent Reasoning, which allocates more recurrent refinement steps in latent space to harder tokens.
Serving is covered by a modified vLLM (v0.26.0 branch) and a Docker image, with licensing under the NVIDIA Open Model License. The mere fact that a model pointing in this direction has been released at a size that runs on a single GPU should make robotics and VLA research that much more approachable.
Jev Sees “Fastest Adoption Ever” on Vercel's AI Gateway — Domestic Experiments Are Starting to Produce Numbers
Jev (TypeSafe AI, in early access since September 15) doesn't write prose; it only returns judgments over structured choices. We've been following it on this blog for several days, and Vercel has now reported on its own blog that “Jev is the fastest-adopted model in AI Gateway history.” Just five days after the announcement, the platform itself is already discussing adoption data — a sign the topic is more than a passing fad.
Hands-on reports are likewise shifting from first impressions to hard numbers. In an experiment shared on Qiita, the model was given 60 fictional sales-call logs with ground-truth labels and asked a five-way classification of “what was the outcome of the call”; it answered 60/60 correctly, including 24 deliberately tricky calls. Batching five questions together yielded a median of 250 milliseconds, and processing all 60 logs reportedly cost about ¥0.6.
Evaluation for hiring and moderation use cases is progressing too. An experiment on Zenn attempted first-pass screening on 76 synthetic resumes and work histories; when “meets the required qualifications” and “advance to interview” were treated as the same judgment, the finding was that candidates with transferable track records tend to slip through. Another evaluation covered 72 post-moderation cases, noting that agreement with ground-truth labels (about 90%) and the share that can be automated without human review (a bit under 30%) are two different things. The more weight you put on not missing anything, the smaller the share that can pass on automated judgment alone — useful material for deciding where to draw the line before adopting.
The range of applications is widening as well. In one example, it was built into a childcare-logging app to predict “around what time is the next feeding,” with time arithmetic left to ordinary code and only the judgment delegated to Jev — a textbook illustration of where a judgment-specialized model fits. Explain articles that organize how it works around the metaphor “if existing LLMs are the CPU, Jev is the GPU” have appeared too, so the explainer side has come together as well.
Hijacking a Real Artist's Spotify with AI Music — An Experiment That Was “Creepily Easy”
A reporter at 404 Media ran an experiment that effectively took over a real artist's Spotify presence using AI-generated music, and reported that it was “creepily easy.” The method — impersonating a real musician and distributing AI-generated tracks under their name — has also become a talking point on Hacker News. It's a case of generative AI's fake-content problem surfacing in a form that directly hits music-streaming revenue and reputation, beyond text and images. Whether platform-side verification and countermeasures are keeping up is likely to come under scrutiny.
Anthropic's IPO May Slip from October to November — Cost Structure Includes a $1.25 Billion-a-Month SpaceX Contract
The Decoder reports that Anthropic, following OpenAI, is postponing its IPO. The listing, expected in October 2026, is now seen sliding to November, apparently with the aim of going public after presenting strong third-quarter results. Investors are pricing in a valuation of around $2 trillion, while infrastructure costs — said to reach $1.25 billion per month on the SpaceX contract alone — plus unresolved security risks are reportedly complicating the listing. Building on yesterday's report on this blog that the company was “preparing to list as early as November,” the reasons for the delay and the cost breakdown have now become concrete.
Japan-Built Japanese Embedding Model Takes First Place on a Search Benchmark
Sionic AI announced that a Japanese embedding model developed through its EUREKA project has taken first place in Japanese search on a retrieval benchmark (as of August 7, 2026). EUREKA is described as an effort to build universal, robust embeddings that can reach knowledge regardless of language, domain, or query format. Because this is a component that directly shapes Japanese RAG quality, a domestic model topping the benchmark adds a practical option.
One “Please Review This” Burns 36.6 Million Tokens — Settings to Rein in Claude Code's Consumption
On Zenn, an incident analysis is drawing attention: asking Claude Code to review a PR consumed roughly 36.6 million tokens (about $40 at API rates, around ¥6,000) for a diff of only about 250 lines. Dozens to hundreds of shell commands reportedly ran, eating through rate limits in an instant. The author traced through the logs to see how the tokens piled up and compiled settings and prompting techniques for running reviews fast and accurately while keeping consumption down. This blog has also covered a breakdown of the structure that loads tens of thousands of tokens at session start — but a single request message can spike consumption at any point, not just at the beginning. Convenience and cost management are two sides of the same coin with agentic tools, and these are real-world numbers worth knowing early in adoption.