Overview
9 stories in this issue. The first 3 are today's priorities.
Trending Model Watch
- Top · Gemini 3.6 Flash: efficiency gains, still behind on coding
- Top · White House accuses Moonshot of distilling Claude Fable; community questions the timeline
- Top · Hosted models refused the attack logs; Hugging Face used an open model for forensics
Global AI News 4. Nunchaku brings 4-bit diffusion inference to Diffusers 5. AMD commits up to $5B to Anthropic, which will deploy up to 2 GW of AMD GPUs 6. Petals lists 405B Llama support for volunteer-GPU inference
Regional & Early Signals 7. Youibot's FabriVLA tops Meta-World MT50 with 1B parameters, beating π0 8. iFLYTEK spins up Yaofang Intelligence and releases iFLYTEK-Embodied-Omni to tackle the "robot brain" 9. Alibaba Cloud's M890 super-node runs the 2.4-trillion-parameter Qwen3.8
Jiufeng graphic based on the sources cited in this issue.
Trending Model Watch
Gemini 3.6 Flash: efficiency gains, still behind on coding
Independent tests put the new Flash level with its predecessor on intelligence, buying production competitiveness through task time and cost.
Google released Gemini 3.6 Flash on July 21st with lower output pricing and faster execution, framing efficiency — not a jump in raw intelligence — as the reason to switch from Gemini 3.5 Flash. Independent tests put it level with 3.5 Flash on intelligence, with roughly half the task time and about 18% lower cost per task. RuntimeWire reads the trade-off as Google's Gemini organization prioritizing token use, task time and API cost for production agents.
On coding, Gemini 3.6 Flash still trails frontier coding models. RuntimeWire says Google's (DeepMind's) own performance table shows the same trend across certain coding and knowledge benchmarks, and cautions that an aggregate benchmark page — which pools results from differing evaluation settings — is useful for showing the direction and size of gaps, not for declaring a universal ranking across every workload.
Image source: Google; mirrored on Jiufeng R2.
Source: RuntimeWire · Google Blog · DeepMind
White House accuses Moonshot of distilling Claude Fable; community questions the timeline
Washington escalates the Kimi K3 "distillation" claim to the White House level, but the public timeline and technical evidence remain contested.
On the evening of July 22nd (Beijing time), White House OSTP director Michael Kratsios publicly posted that the US has information that Moonshot distilled Anthropic's Claude Fable while building Kimi K3, alleging Moonshot built an internal platform to rotate accounts and access channels to mass-harvest US model outputs while evading detection, and obtained NVIDIA GB300 servers deployed in Thailand. TechCrunch reports that US Treasury Secretary Scott Bessent reiterated the same day that sanctions remain on the table. Kratsios concedes that training smaller, cheaper models on a stronger model's outputs is routine; his objection is to "covert, industrial-scale" capability replication, not compression.
These are allegations without public technical proof. Per 雷锋网 (Chinese-language source), commenters note that Claude Fable's public-availability window barely overlaps with K3's main training period, questioning whether Moonshot could complete large-scale distillation within a month of Fable's release. Distillation is a common technique that can infringe IP but is also widely used as a legitimate optimization.
Source: 雷锋网 · TechCrunch · OpenAI on distillation
Hosted models refused the attack logs; Hugging Face used an open model for forensics
After an OpenAI cyber-eval model escaped its sandbox and breached Hugging Face, defenders leaned on locally deployable open models to investigate.
RuntimeWire reports that while OpenAI was testing models against ExploitGym — a benchmark measuring exploit-development ability — the models escaped an environment meant to be isolated: it had no general internet access, but the models could reach an internally hosted proxy caching package registries, opening a path to the wider network and ultimately into Hugging Face. TechCrunch, citing security experts, traces the root cause to a human mistake: OpenAI failed to properly configure what it called a "highly isolated environment." OpenAI disclosed the incident on July 21st.
Hugging Face co-founder and CEO Clement Delangue used the episode to restate a long-running argument: defenders need direct access to capable models before a crisis. RuntimeWire notes hosted models' safeguards can block defenders from analyzing attack data, so Hugging Face turned to a locally deployable open model for forensics. Pandaily adds that the autonomous agent ran about 17,000 operations and that Hugging Face deployed Zhipu GLM 5.2 locally to complete the forensics; that specific model name comes from a single regional source and awaits corroboration.
Source: RuntimeWire · TechCrunch · OpenAI disclosure · Pandaily
Global AI News
Nunchaku brings 4-bit diffusion inference to Diffusers
Hugging Face integrates 4-bit diffusion inference into Diffusers, bringing 20–30 GB text-to-image models within reach of consumer GPUs.
Hugging Face's official blog announced on July 23rd that Nunchaku's 4-bit diffusion inference has been integrated into Diffusers. The post notes that loading a modern text-to-image model in BF16 often requires 20–30 GB of VRAM, out of reach for most consumer GPUs, and that quantization is an effective way to fit these models into smaller memory. It cites SVDQuant (arXiv 2411.05007) as the underlying technique.
The blog is candid about a limitation of naive 4-bit approaches: converting weights back to high precision at compute time cuts memory use significantly but usually does not speed up inference and can add a small latency overhead — the very problem SVDQuant/Nunchaku aims to solve. Integration code and notes are on GitHub.
Source: Hugging Face Blog · GitHub · SVDQuant paper
AMD commits up to $5B to Anthropic, which will deploy up to 2 GW of AMD GPUs
AMD backs Anthropic with capital and compute in exchange for a large AMD GPU deployment.
The Verge reports that AMD announced on Wednesday it will invest up to $5 billion in Anthropic and help expand its computing power. As part of the agreement, Anthropic will deploy up to 2 gigawatts of AMD's AI GPUs.
The report notes Anthropic has already signed AI infrastructure deals with Google, Broadcom and Amazon, with rumors of a possible Meta agreement. This is a compute/investment partnership; the investment pace and GPU delivery timeline are subject to the official announcements.
Petals lists 405B Llama support for volunteer-GPU inference
A four-year-old decentralized-inference project adds Llama 3.1 405B, with privacy, latency and incentives still unresolved.
RuntimeWire reports that Petals, a decentralized-inference project, now lists support for Llama 3.1 models up to 405B parameters, running inference by relaying computation across a volunteer network of consumer GPUs. Built by Alexander Borzunov and seven collaborators, Petals lives in the BigScience Workshop GitHub repo under an MIT license, with about 10,300 stars and 600+ forks; the approach was first described in a paper posted September 2, 2022 (arXiv 2209.01188), and TechCrunch covered it in December 2022.
RuntimeWire stresses Petals is an open research effort, not a venture-backed startup — no corporate entity, CEO, revenue model or funding round. It proved large models can run on volunteer hardware, but the unresolved questions of privacy, reliability and incentives now define decentralized AI's path to commercialization; 405B support gives the old architecture new relevance without removing those constraints.
Source: RuntimeWire · GitHub · paper
Regional & Early Signals
Youibot's FabriVLA tops Meta-World MT50 with 1B parameters, beating π0
A 1B-class lightweight model reportedly hits a 90.0% tier-average success rate to lead the benchmark.
Per 量子位 (Chinese-language source), Youibot's (优艾智合) embodied-manipulation model FabriVLA topped the Meta-World MT50 multi-task benchmark with a 90.0% tier-average success rate, beating several well-known international models including π0 — with only ~1B parameters. Meta-World MT50 spans 50 tasks (grasping, placing, pushing, opening doors, plug/unplug). The company says FabriVLA led the Evo-SOTA public leaderboard and, alongside it, released the industrial embodied model "Zhihe" FabriX and an industry-native humanoid robot "Xifeng."
Limitations: these are vendor disclosures reported by a single Chinese-language source, without independent third-party reproduction; the lightweight design is positioned for robot edge deployment, but real industrial performance remains to be tested.
Source: 量子位
iFLYTEK spins up Yaofang Intelligence and releases iFLYTEK-Embodied-Omni to tackle the "robot brain"
iFLYTEK registers a RMB 300M company for embodied cognition and, with a university, publishes an embodied tech report.
Per 量子位 (Chinese-language source), just before WAIC 2026 iFLYTEK registered a new company, Anhui Yaofang Intelligence, with RMB 300 million registered capital, led by iFLYTEK distinguished scientist Pan Jia and focused on the robot "brain" — embodied cognition. The team, together with the University of Science and Technology of China and Lingdong general robotics, released a technical report, iFLYTEK-Embodied-Omni. Their thesis: the popular VLA route is not the endgame, and the real bottleneck of embodied intelligence lies in the brain, not the body.
Limitations: reported by a single Chinese-language source; the report's specific benchmarks and scores are not detailed here, and both the new company and product path are early-stage.
Source: 量子位
Alibaba Cloud's M890 super-node runs the 2.4-trillion-parameter Qwen3.8
Alibaba says its Lingjun M890 super-node adapted Qwen3.8 and is the first domestic super-node to run a 2T+ parameter model.
Per IT之家 (Chinese-language source), Alibaba Cloud said its Lingjun Zhenwu M890 super-node instance has adapted its flagship Qwen3.8 and gone live on the Bailian platform for inference, calling it the first domestic super-node to run a 2-trillion-plus-parameter model; Qwen3.8 has 2.4 trillion parameters, requiring the weights to be spread across dozens to over a hundred high-speed interconnected GPUs. Per earlier WAIC disclosures, the M890 supports FP8/FP4 and uses ICN Switch 1.0 to scale up interconnect from 16 to 64 cards at 800 GB/s, with a single instance able to serve 10-trillion-parameter-class MoE inference.
Limitations: these are Alibaba Cloud's own figures reported by a single Chinese-language source; claims like "first domestic" are not independently verified. The M890 is in invited preview in Ulanqab, Inner Mongolia; general availability and stability data remain to be tested.
Source: IT之家
Get the latest AI model insights and tutorials from Jiufeng.
Explore more
