Updates

All updates

• 5 October 2026
  Reflex-1: the latest public weights, reserved evaluation results and CPU measurements are now on the model page. 421M parameters; 2.17 GiB peak process memory in the recorded CPU replay.
• 5 October 2026
  Samsara-VLA and Samsara-VLA Tiny: watch eight recorded robot tasks across four LIBERO suites. Compare the 217M and 137M policies, their success rates and local inference options.
• 2 October 2026
  reflex-1 is available on Hugging Face with public FP32 weights, tokenizers and Transformers inference code. The model page includes dated game and browser recordings, evaluation conditions and resource measurements.
• 2 October 2026
  Samsara VLA: published all forty task counts and a CSV of 10,000 recorded trials across five checkpoints, with architecture and training-resource details.
• 2 October 2026
  re1: explore the research preview through recorded examples of video recall and captured environments.
• 2 October 2026
  Colony: persistent personal agents, scheduled work, voice, Mac computer use and small apps. Read the current technical overview.
• 1 October 2026
  Announced reflex-1 and Samsara VLA. Detailed model pages include latency, accuracy and resource measurements.
• 29 September 2026
  reflex-1 H200 test: 27.5 ms median and 30.5 ms p95 across 240 recorded requests, using FP32 and loopback HTTP.
• 24 September 2026
  reflex-1 M4 CPU test: 345.2 ms median for Banking77 intent selection, using FP32 and one thread.
• 17 September 2026
  Samsara VLA completed 1,897 of 2,000 LIBERO episodes. The selected checkpoint scored 94.85% overall and 99.6% on Object.
• 19 August 2026
  Open roles in robot learning, 3D vision, agent systems and deployment engineering.
• 18 August 2026
  Journal: local inference and hardware requirements.
• 12 August 2026
  Four robot-grading protocols, with 400-episode results across three historical checkpoints.
• 5 August 2026

  Added a plain-text version of each page at llms.txt.

• 14 July 2026
  Journal: development evaluations and reserved confirmation sets.
• 6 May 2026
  Journal: deploying models inside a local application.
• 11 February 2026
  Journal: the lab's reasons for working on local models.
• 11 March 2025
  Research notes on model weights and adaptation.