CEO of AutoTrust. Inference is Training and Model Size is Dynamic.
- Tweets23
- Following30
- Followers3
- Likes72
ALT AutoTrust AI's JEV-27B-VL just hit #5 globally on @huggingface Trending, out of 3.1M+ models. GEV-26B-Decide is #9. Two open decision models in the global Top 10. 1.7M+ downloads combined. JEV-27B-VL: calibrated visual decisions for agents. GEV-26B-Decide: adaptive thinking. Fast when confident, deeper when uncertain. Agents need models that decide, not just describe. Builders are choosing ours. Thank you to everyone building with us. huggingface.co/autotrust
ALT AutoTrust AI decision models just passed 2M downloads on @huggingface. In Model Pulse's "New this month" chart (week ending Oct 5), three AutoTrust models rank in the top 15: #3 JEV-27B-VL: 1.3M #7 GEV-26B-Decide: 446.5K #13 JEV-9B: 305.2K That's a clear signal: builders are making AutoTrust a preferred open-weights provider for decision models. Why? Agents need a different interface: visual state in → calibrated action probabilities out JEV-27B-VL doesn't just describe an image. Give it the image state, the task context, and the candidate actions, and it returns a probability for every available next move. Your agent knows what to do next, and how confident to be. Thank you to the open-source builders exploring these models with us. We're just getting started. Model family: https://huggingface.co/autotrust JEV-27B-VL: https://huggingface.co/autotrust/JEV-27B-VL Source: https://huggingface.co/spaces/tardellirs/model-pulse
ALT Hugging Face Trending Update Two AutoTrust AI models are now in the global Top 5: #2 — JEV-27B-VL — 1.53M downloads #5 — GEV-26B-Decide — 855K downloads Together with JEV-9B, our latest open decision-model releases have now passed 2.7M downloads. JEV-27B-VL turns visual state into calibrated next-action probabilities. GEV-26B-Decide combines fast System 1 decisions with adaptive System 2 reasoning when uncertainty makes deeper thinking worthwhile. From visual agents and computer use to tool routing, agent evaluation, and research workflows, developers are building with a different interface: state in → calibrated decisions out → action Thank you to everyone downloading, testing, and contributing feedback. JEV-27B-VL: https://huggingface.co/autotrust/JEV-27B-VL GEV-26B-Decide: https://huggingface.co/autotrust/GEV-26B-Decide All AutoTrust AI models: https://huggingface.co/autotrust
ALT 72.24 minutes to GPT-2 on 8×H100. AutoTrust's ScienceGuru, running Guru Turbo 1.2, posted the fastest result we've found on Andrej Karpathy's Time-to-GPT-2 benchmark (one run, self-reported): 27% under the official record and 9.6 min ahead of the best community recipe. The field Reported times on the official leaderboard and in public nanochat PRs, as of Sept 24: - 72.24 min · ScienceGuru (AutoTrust) · 1 run - 81.84 min · Giovanni Zinzi · 6 runs - 91.74 min · Oriole Networks · 3 runs - 94.58 min · Martin Jurča (Seznam.cz) · 6 runs - 94.6 min · Weco-optimized run · 3 runs - ~99 min · Official record, Karpathy's autoresearch round 2 · 5 runs Zinzi also reported a 73.92-min experiment with a Cosmopedia data mix (3 runs), which he set aside after it regressed at a smaller model size. The lead - 27% faster than the official record (−26.8 min) - 12% faster than the best community recipe (−9.6 min) - 19–22 min ahead of Oriole Networks, Seznam.cz and the Weco-optimized run How Starting fr
ALT 72.24 minutes to GPT-2 on 8×H100. AutoTrust's ScienceGuru, running Guru Turbo 1.2, posted the fastest result we've found on Andrej Karpathy's Time-to-GPT-2 benchmark (one run, self-reported): 27% under the official record and 9.6 min ahead of the best community recipe. The field Reported times on the official leaderboard and in public nanochat PRs, as of Sept 24: - 72.24 min · ScienceGuru (AutoTrust) · 1 run - 81.84 min · Giovanni Zinzi · 6 runs - 91.74 min · Oriole Networks · 3 runs - 94.58 min · Martin Jurča (Seznam.cz) · 6 runs - 94.6 min · Weco-optimized run · 3 runs - ~99 min · Official record, Karpathy's autoresearch round 2 · 5 runs Zinzi also reported a 73.92-min experiment with a Cosmopedia data mix (3 runs), which he set aside after it regressed at a smaller model size. The lead - 27% faster than the official record (−26.8 min) - 12% faster than the best community recipe (−9.6 min) - 19–22 min ahead of Oriole Networks, Seznam.cz and the Weco-optimized run How Starting fr
ALT 72.24 minutes to GPT-2 on 8×H100. AutoTrust's ScienceGuru, running Guru Turbo 1.2, posted the fastest result we've found on Andrej Karpathy's Time-to-GPT-2 benchmark (one run, self-reported): 27% under the official record and 9.6 min ahead of the best community recipe. The field Reported times on the official leaderboard and in public nanochat PRs, as of Sept 24: - 72.24 min · ScienceGuru (AutoTrust) · 1 run - 81.84 min · Giovanni Zinzi · 6 runs - 91.74 min · Oriole Networks · 3 runs - 94.58 min · Martin Jurča (Seznam.cz) · 6 runs - 94.6 min · Weco-optimized run · 3 runs - ~99 min · Official record, Karpathy's autoresearch round 2 · 5 runs Zinzi also reported a 73.92-min experiment with a Cosmopedia data mix (3 runs), which he set aside after it regressed at a smaller model size. The lead - 27% faster than the official record (−26.8 min) - 12% faster than the best community recipe (−9.6 min) - 19–22 min ahead of Oriole Networks, Seznam.cz and the Weco-optimized run How Starting fr
