Decompute is an artificial intelligence (AI) startup focused on decentralizing AI development and infrastructure.decompute.run Cupertino, CAJoined September 2024
Download Decompute Claude Gateway, to save on tokens, and also enable your personal model to learn from your interactions/serve you offline.
claude.decompute.run 🚀
Claude Fable 5 will be available again globally tomorrow.
After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding
Companies want AI to learn from sensitive work without exposing the underlying data.
Echelon trains across privacy boundaries with aggregate-only updates.
Decompute: train without centralizing. Tune without retraining. Deploy without lock-in.
lnkd.in/gBt2DEzT
The most important thing about Decompute Gateway isn’t that it saves tokens. It’s that it creates the control point where Claude usage can become a learning system for local AI Infrastructure for enterprises and professionals.
We guarantee your Claude costs drop. If they don’t, your trial gets extended.
Decompute Gateway: from cutting tokens → owning your AI.
One-click VS Code install. Zero data sent to us. Runs local.
claude.decompute.run
Claude resends the same context on every request and you end up paying frontier prices for it. Decompute Gateway runs locally between Claude Code and Anthropic. A local model learns which context matters and compresses the rest before it hits your bill.
Cut Claude Code tokens before they hit your bill for free 💵
84,000 raw tokens
↓
Decompute Claude Gateway
-78% on this request
↓
18,700 sent tokens
Same Claude Code. Same workflow. Fewer tokens. 🧮
claude.decompute.run
Nebula-S-V2 delivers frontier-class reasoning at 3B parameters: it beats MAI-Thinking-1 31B/1T parameters & Gemma 4 31B on GPQA Diamond, stays within ~2 points of their MMLU-Pro scores, and outperforms the closest active-parameter Gemma 4 MoE on the overlapping public benchmarks.
I have been developing my own VLA for pick and place tasks. The model now has intuition to anticipate potential failures during an episode and adjust grip or other relevant factors in its attempts based on the internal signal. Fun times.
Nebula-S-V2 delivers frontier-class reasoning at 3B parameters: it beats MAI-Thinking-1 31B/1T parameters & Gemma 4 31B on GPQA Diamond, stays within ~2 points of their MMLU-Pro scores, and outperforms the closest active-parameter Gemma 4 MoE on the overlapping public benchmarks.
Presenting Echelon, with 2,139-2,176 tokens/s across evaluated WAN and non-IID treatments, with zero data leaving your devices. This is the future for enabling training across heterogeneous clusters.
Our paper is now live: arxiv.org/abs/2606.02958
We quietly shipped self-improving
on-device agent last year. You can download it here now: decompute.run/blackbird/down…
Zero data leaves your device. 🪴
Most distributed/federated stacks assume cross-site model exchange, then retrofit secure agg/DP/TEEs.
That’s why compliance reviews are brittle.
Echelon starts with a hard rule: device-level model state never leaves a boundary.
📌My latest substack is now live. Link in comments.
🚀 Echelon is bending the aggregate-only training frontier.
In 1B LoRA adaptation, we hit 3.887 val loss / 48.75 PPL, best vs tuned DiLoCo-family baselines, while sustaining 2,139 to 2,176 tok/s under WAN + non-IID stress. At 100ms WAN: 95 min to target vs 145 for DiLoCo+SA.
20 Followers 113 Followingbug-bitten dev building AI that works, not hallucinates
Building Achaia Labs @achaialabs
☕ https://t.co/tX4OPtijbx • https://t.co/kewkRwvFmj
100 Followers 301 Followinghttps://t.co/FcXg2lJm2Z puts AI in your Org Chart with Autonomous Agents. Build an AI-Native Company with https://t.co/FcXg2lJm2Z
22 Followers 160 FollowingA system that understands your work, your goals, and your context—so you can focus, execute, and grow without the constant overwhelm.
925K Followers 6K FollowingPresident & CEO @ycombinator —Founder @garryslist—Creator of GStack & GBrain—designer/engineer who helps founders—SF Dem accelerating the boom loop
310K Followers 1K FollowingBuilding new things @thinkymachines. Also dabble in robotics at NYU. Cofounded @PyTorch. AI is delicious when it is accessible and open-source.
78K Followers 1K FollowingCo-founder and CTO at @huggingface 🤗. ML/AI for everyone, building products to propel communities fwd. @Stanford + @Polytechnique