English text is being prepared.
LatestARC Prize publishes Meta Muse Spark Open's evaluation results on ARC-AGI
Why read thisARC Prize gives Meta's new model's scores on three versions of ARC-AGI along with per-task cost, directly comparable with competitors scoring the same.
NVIDIA announced a $3 billion investment in Meta's equity offering, and the two companies will expand their collaboration on the personal agent product Muse and the Muse series of models to improve model inference and training performance. The investment is part of Meta's $95 billion fundraising plan, and AMD and Broadcom have also expressed interest in investing. Meta expects full-year capital expenditure of $135 billion to $145 billion and will consider adopting NVIDIA's OpenShell secure agent reference architecture and expanding deployments of infrastructure including Vera Rubin, Vera CPU, and Spectrum-X networking.
Why read thisNVIDIA takes a $3 billion stake in Meta and ties in a Muse agent collaboration, giving readers a clear view of how the compute layer and the application layer are betting on each other.
Google and Meta open-sourced TorchTPU at PyTorch Conference North America 2026. This is a native PyTorch backend for Google TPUs, replacing PyTorch/XLA from 2020; migrating a model only requires changing model.to("cuda") to model.to("tpu"). The official release also open-sources four repositories: TorchTPU, vLLM-TorchTPU, SGLang-TorchTPU, and TorchTitan. Hugging Face models can run directly, with native support from the vLLM and SGLang inference engines. Google's published results include running the Qwen3 235B MoE model on 8192 TPU v7x chips and reaching 28% model FLOPs utilization, but the announcement also states that a one-line change only gets the model running; achieving peak performance still requires adjusting shapes, sharding, and operators.
Why read thisThe original text explains how TorchTPU is integrated, the repositories already open-sourced, and the measured scale, and also notes that a one-line change is only a starting point and tuning still requires investment.
Anthropic announced the launch of AI agent Claude Duet and released a research preview, available in standalone apps on iOS and Android and on the claude.ai website, and already included in the Pro and Max plans. The product is built on the Opus 5.5 model and targets small and medium businesses and professionals; it can send daily briefings, monitor connected apps, and perform tasks such as building dashboards, scheduling meetings, and creating financial forecasts. It is equipped with a virtual machine with an independent browser, can keep working while the user is away, and seeks consent before acting on the user's behalf. Anthropic notes that the product is still at an early stage and may misread conversation content or fail to accurately judge urgent matters, and performs somewhat poorly on personal tasks such as buying tickets.
Why read thisAnthropic turns the agent into a standalone app and integrates it into the Pro and Max plans, which readers can use to compare the positioning differences of similar assistants from Meta and OpenAI.
Fireworks launched Meta's open-weight model Muse Spark Open on the day of release. It is an 816B-parameter MoE model that activates 42B parameters per token, with a 1M-token context window, targeting long-horizon coding and agentic tasks. Fireworks says it ranks first among 75 models on Vals AI's Harvey Legal Agent Benchmark leaderboard with 25.4%, nearly 4 times that of Claude Opus 5 (6.7%); on the Rogo Big Finance Bench it surpasses Claude Opus 5.5 at every difficulty level. On Fireworks, the model has zero data retention enabled by default, uses only US-based compute at no extra charge, and supports fine-tuning with serverless LoRA.
Why read thisThe original article lists the model's specific scores and comparisons on multiple professional benchmarks, allowing readers to judge where open-weight models stand on regulated professional tasks.
The SGLang team, in collaboration with Google and Meta, has open-sourced SGLang-TorchTPU, a PyTorch-based native backend that brings the full SGLang serving stack to Google Cloud TPU and is already in large-scale production use at Meta. Initial support covers models including Muse Spark Open, GLM-5.3, Kimi K3, DeepSeek V4, Qwen3.8, and Gemma 4, with features such as speculative decoding, KV cache offloading, PD disaggregation, and FP8 KV cache. The team also used TorchTitan and Miles to complete GRPO training of Qwen3-4B on a single 4-chip TPU v7x, with the reward curve matching a 6× H200 baseline within random fluctuation.
Why read thisThe original post details the plugin architecture, supported models, and CI coverage, which can be used to assess the deployment boundaries of running SGLang on TPU.
Google and Meta's PyTorch teams have collaborated to launch the open-source project TorchTPU, providing a native PyTorch backend for Google TPUs; migration only requires changing model.to("cuda") to model.to("tpu"). The project supports both eager and torch.compile modes, adds a new DeferAndFuse eager mode, covers about 450 Aten operators, and has been verified to run out of the box on 4,917 Hugging Face models, with a Tier 1 pass rate of 98% and Tier 2 numerical consistency relative to GPUs of 97%. According to the official announcement, it has already been used in the inference and training production workloads of some major Google Cloud customers, and is integrated with vLLM, SGLang, TorchTitan, Ray, and TorchRec.
Why read thisThe original article presents TorchTPU's design principles, architectural details, and multiple sets of measured data, which can be used to assess the actual cost of migrating PyTorch workloads to TPUs.
Why read thisMeta has published the technical report for a 30B open-weight model, covering architecture, distillation recipe, and quantization, for readers interested in on-device agent deployment.
Meta Superintelligence Labs releases its second-generation model family Muse Wonder and Muse Wonder Full; the former is positioned as a fast frontier model, the latter as the full version, and Meta says they target personal and professional agent work. On Artificial Analysis's Intelligence Index v4.3, Wonder Full scores 56 and Wonder scores 52, higher than Muse Spark 1.3's 48, while Opus 5.5 leads with 58. Specifications are 1M context, input supports text, images, video and audio, output is text only, and pricing is 4 USD per million tokens for input, 0.20 USD for cached input, and 20 USD for output.
Why read thisSummarizes third-party evaluations and pricing, allowing comparison against models such as Opus 5.5 and GPT-6 to judge the actual position of Meta's new models.