←Search results

Search “Google”

By relevance · Total 1215 results Clear search

Thursday

  1. FactRelated story

    Google open-sources TorchTPU native PyTorch stack

    Google collaborated with Meta's PyTorch team to launch the open-source project TorchTPU, providing a native PyTorch backend for Google TPU; 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 GPU of 97%. The official statement says it has been used in the inference and training production workloads of some Google Cloud major customers, and is integrated with vLLM, SGLang, TorchTitan, Ray, and TorchRec. Subsequent reports said that at PyTorch Conference North America 2026, Google open-sourced TorchTPU, whose runtime no longer depends on JAX, and existing workloads can be migrated with only minor code changes; PyTorch noted that TorchTPU's default Skinny Collectives scheme suffers roughly 30–40% end-to-end performance loss under distributed workloads compared with optimized schemes, and achieving peak performance still requires Pallas kernels and vLLM-T

    Matched fieldsTitle

    View story:Google open-sources TorchTPU native PyTorch stack

Sunday

  1. Story1 reports · 1 facts

    Google releases TorchTitan branch adapted for TPUs

    On October 20, 2026, Google released a TorchTitan branch adapted for Google Cloud TPUs. According to the Google Developers Blog, Google integrated TorchTPU into Meta's TorchTitan, and Cloud TPUs can now directly train PyTorch models through the native backend torch.device("tpu") without rewriting them into JAX. The TPU branch provides pretraining recipes for Llama 3 70B/405B, Qwen 3 8B/32B/235B MoE, and FLUX.1 12B, and replaces standard implementations with Pallas kernels such as Splash Attention. On FLUX, SimpleFSDP enables TorchTitan to reach performance parity of 99% with JAX MaxDiffusion. There are currently no subsequent development facts.

    On October 20, 2026, Google released a TorchTitan branch adapted for Google Cloud TPUs. According to the Google Developers Blog, Google integrated TorchTPU into Meta's TorchTitan, and Cloud TPUs can now directly train PyTorch models through the native backend torch.device("tpu") without rewriting them into JAX. The TPU branch provides pretraining recipes for Llama 3 70B/405B, Qwen 3 8B/32B/235B MoE, and FLUX.1 12B, and replaces standard implementations with Pallas kernels such as Splash Attentio

    Matched fieldsTitle

  2. Story1 reports · 1 facts

    Princeton and Google paper finds 4B model trained on agent states achieves higher task success rate

    Princeton University and Google published a paper studying the performance of a 4B model trained on agent states in agent tasks. The paper reports that the most accurate decision model on isolated problems solved only 69.5% of agent tasks, while the 4B model trained on agent states solved 91.5%. Its harness delegates tool selection at each step to a fast decision model, adopting its answer only when confidence exceeds a threshold fitted on the agent's own runs, and trained AgentDecision-4B on the states the agent actually reached, including states after mistakes. On tasks similar to the training data, this type of 4B model reduced agent costs by 81.1% to 89.0% and increased speed by up to 2.5 times, with a success rate gap of within 3 percentage points; delegating all tool selections without a threshold lowered the success rate by 19.0 percentage points. The paper is available at arxiv.org/abs/2610.21412.

    Princeton University and Google published a paper studying the performance of a 4B model trained on agent states in agent tasks. The paper reports that the most accurate decision model on isolated problems solved only 69.5% of agent tasks, while the 4B model trained on agent states solved 91.5%. Its harness delegates tool selection at each step to a fast decision model, adopting its answer only when confidence exceeds a threshold fitted on the agent's own runs, and trained AgentDecision-4B on th

    Matched fieldsTitle

  3. Story1 reports · 1 facts

    Google Research releases the Harness2 framework

    Google Research has released the Harness2 framework, which lets agents improve their own harness while the model remains frozen. According to the same report, Gemini 3.5 Flash improved by 8.4 points on WorkBuddy-Bench and 7.5 points on JobBench, and improved on all 3 benchmarks; Skills and scripts accounted for about 50% of the edits, and removing skills would cost 5.3 points. This news currently comes from a single-source report, with no subsequent independent developments yet.

    Matched fieldsTitle

Tuesday

  1. FactRelated story

    Google releases TorchTitan branch adapted for TPUs

    On October 20, 2026, Google released a TorchTitan branch adapted for Google Cloud TPUs. According to the Google Developers Blog, Google integrated TorchTPU into Meta's TorchTitan, and Cloud TPUs can now directly train PyTorch models through the native backend torch.device("tpu") without rewriting them into JAX. The TPU branch provides pretraining recipes for Llama 3 70B/405B, Qwen 3 8B/32B/235B MoE, and FLUX.1 12B, and replaces standard implementations with Pallas kernels such as Splash Attention. On FLUX, SimpleFSDP enables TorchTitan to reach performance parity of 99% with JAX MaxDiffusion. There are currently no subsequent development facts.

    Matched fieldsTitle

    View story:Google releases TorchTitan branch adapted for TPUs

Wednesday

Sunday

Friday

Monday

  1. FactRelated story

    Princeton and Google paper finds 4B model trained on agent states achieves higher task success rate

    Princeton University and Google published a paper studying the performance of a 4B model trained on agent states in agent tasks. The paper reports that the most accurate decision model on isolated problems solved only 69.5% of agent tasks, while the 4B model trained on agent states solved 91.5%. Its harness delegates tool selection at each step to a fast decision model, adopting its answer only when confidence exceeds a threshold fitted on the agent's own runs, and trained AgentDecision-4B on the states the agent actually reached, including states after mistakes. On tasks similar to the training data, this type of 4B model reduced agent costs by 81.1% to 89.0% and increased speed by up to 2.5 times, with a success rate gap of within 3 percentage points; delegating all tool selections without a threshold lowered the success rate by 19.0 percentage points. The paper is available at arxiv.org/abs/2610.21412.

    Matched fieldsTitle

    View story:Princeton and Google paper finds 4B model trained on agent states achieves higher task success rate

  2. FactRelated story

    Google Research releases Harness² framework

    Google Research has released the Harness2 framework, which lets agents improve their own harness while the model remains frozen. According to the same report, Gemini 3.5 Flash improved by 8.4 points on WorkBuddy-Bench and 7.5 points on JobBench, and improved on all 3 benchmarks; Skills and scripts accounted for about 50% of the edits, and removing skills would cost 5.3 points. This news currently comes from a single-source report, with no subsequent independent developments yet.

    Matched fieldsTitle

    View story:Google Research releases the Harness2 framework