Decompute is an artificial intelligence (AI) startup focused on decentralizing AI development and infrastructure.
Joined September 2024
- Tweets 208
- Following 15
- Followers 1,843
- Likes 433
49 Photos and videos
Pinned Tweet
13 Jun 2025
BlackBird is now on Windows!
Spin up powerful AI agents in just a click.
Train models locally, privately, and offline, right from your laptop.
Download now: decompute.run/releases
The full power of AI is now in your hands.
40
79
694
2,556,534
Jun 12
Cut Claude Code tokens before they hit your bill for free đź’µ
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
29,912
Jun 8
15K Downloads for Nebula-S-V1. Can’t wait to ship V2! ✌️
Jun 4
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.
160
Decompute AI retweeted
Jun 5
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.
1
1
88
Decompute AI retweeted
Jun 4
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.
1
4
302
Jun 3
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
1
1
83
Decompute AI retweeted
May 29
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. 🪴
1
1
73
Decompute AI retweeted
Jun 1
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.
đź“ŚMy latest substack is now live. Link in comments.
ALT Schema audit across loopback/emulated WAN/real cross-region: per_device_payload = 0B; 0 violations; allowed schemas verified.
1
2
2
103
Decompute AI retweeted
May 28
You can now run our Nebula-S-V1 on @FeatherlessAI
AGI starts here đź§
featherless.ai/models/decomp…
1
2
4
304
May 27
Excited to share our research: decompute.run/research/paper…
May 27
🚀 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.
1
229
Decompute AI retweeted
May 24
🚀 Nebula-S-SVMS2-3B is out.
3B params. Frontier reasoning.
• GPQA Diamond: 86.85
• HMMT Nov 2025: 80.00
• GSM8K: 93.78
• MMLU-Pro: 83.00
Outperforms GPT-OSS-120B (40x larger) and Qwen3-Next-80B-A3B (27x) on GPQA.
Frontier quality, a fraction of the size.
1
1
3
176
Decompute AI retweeted
May 24
Introducing Echelon, @DecomputeAI and team’s work to turn heterogeneous compute clusters into an auditable AI training fabric for Enterprise multi-agents. Our research paper drops soon.
open.substack.com/pub/hinadi…
1
1
3
135
Decompute AI retweeted
May 22
Robots don’t just need a brain. They need a brake pedal.
Most of embodied AI research answers:
“What action should the robot take next?”
Deployment adds a second question:
“Should the robot be allowed to take that action at all?”. Lets chat, if that sounds interesting to you.
1
4
82
Decompute AI retweeted
May 24
The first era of AI infrastructure was built around a clean assumption: put homogeneous accelerators in one place, connect them with fast links, centralize the data, and scale the model from there.
Today it does not match how most enterprises actually operate.
2
1
2
129
Decompute AI retweeted
May 21
In just a month, Nebula-S-v1 is the #1 most-downloaded model on the Hugging Face Hub for math reasoning and competitive math — 10K downloads and counting. Our more advanced reasoning model Nebula-S-SVMS2-3B, will be available soon. Links in comments.
2
1
3
110
May 11
We’re sharing Nebula-S-SVMS2-3B-Internal, an internal evaluation release from Decompute.
Internal evals for this release report:
GPQA: 86.85
HMMT Nov 2025: 80.00
GSM8K: 92.00
MMLU-Pro: 83.00
Model: lnkd.in/gterpHN4
1
4
147
Apr 22
We gave the same messy operating task to an intern and to Nebula-S: customer interviews support tickets partial docs conflicting notes
→produce a decision memo
The intern took hours. Nebula-S-4B produced a strong first pass in minutes. ~15Ă— faster time-to-first-draft.
1
2
147
Apr 21
RT @hinadixit: Nebula-S-v1 just crossed 4,600 downloads. People do not just want bigger models. They want better reasoning. 🚀 https://t.co…
1
Apr 16
RT @hinadixit: 🚀 Meet Nebula-S-SVMS1-4B beating Qwen models in reasoning by 25%, now live on @huggingface.
Built with our patented reasoni…
3
Apr 10
RT @hinadixit: Only top 1% people in AI would know what’s happening here. Comment if you know.
#AI #SLMs #Model #LLM #ArtificialInteligen…
1