c0mpute (@c0mputeAI)

Pinned Post

c0mpute

@c0mputeAI

Jun 4

AI powered by people, not datacenters.

1:04

97

65

559

179,040

c0mpute reposted

leyten

@leyten

17h

Many people don't know that 75% of $ZERO stakers auto-compound their rewards. To bring more attention to this I've added 3 more charts to the data page. A total 3.21M $ZERO has been bought back so far by stakers that auto-compound. data.c0mpute.ai

30

40

151

9,689

c0mpute

@c0mputeAI

22h

The c0mpute roadmap is live! With this roadmap we want to clarify what we're building for and how far along we are. You can check progress live. c0mpute.ai/roadmap

50

71

268

31,785

c0mpute

@c0mputeAI

Jul 3

7/ Full report, benchmark code, and the signed receipts behind every figure: DOI: doi.org/10.5281/zenodo.21178… Code: github.com/leyten/shard

Sharded Inference of a 229B-Parameter MoE over the Public Internet at Interactive Speed

We serve MiniMax-M2.5, a 229B-parameter mixture-of-experts model, split across five consumer RTX 5090s in five European countries. The stages share nothing but the public internet. Single-stream...

zenodo.org

5

14

58

7,403

c0mpute

@c0mputeAI

Jul 3

6/ We ran a test where we kill -9 the coordinator mid-generation, four speculative chunks in flight. A replacement picked up the same warm ring and finished a full request 17 seconds later, without reloading any of the 115GB of weights.

2

7

48

7,958

c0mpute

@c0mputeAI

Jul 3

5/ Every stage signs a hash chain over the computation it actually ran. A node that skips layers fails the request and doesn't get paid. We measured what this protection costs: 0.05ms on an 11.7ms stage span. About 0.4%.

1

24

1,619

c0mpute

@c0mputeAI

Jul 3

4/ Something we didn't expect: all five GPUs benchmark identically, yet some stages ran the same layers 4x slower. It came down to the host CPU and how loaded the box was. A forward pass launches hundreds of small kernels, and slow CPUs launch them slowly.

1

25

1,695

c0mpute

@c0mputeAI

Jul 3

3/ We tried the obvious latency fix, keeping several speculative chunks in flight, and it dies on reasoning: one rejected token throws away everything behind it. You need ~80% per-token acceptance before it pays off. The best drafter we know of gets 74% on novel text.

1

31

2,357

c0mpute

@c0mputeAI

Jul 3

2/ Each decode step travels through all five machines and back before the next one can start. Over real internet links that round trip costs ~350ms, so plain decoding tops out near 5 tok/s. We measured exactly that, flat, on every workload we tried.

1

39

2,854

c0mpute

@c0mputeAI

Jul 3

1/ We published our first technical report today. We ran a 229B model split across five consumer GPUs in five countries over the public internet and measured 12.6 tok/s interactive, 194 tok/s batched. With cryptographic receipts on every request. doi.org/10.5281/zenodo.21178…

33

52

279

77,125

c0mpute reposted

leyten

@leyten

Jun 29

gm, shard is looking good, check the github ;) also releasing an extensive roadmap this week should give you guys an idea of where we’re headed

24

32

165

25,251

c0mpute reposted

leyten

@leyten

Jun 27

The bottleneck of local ai is the ability of running frontier models on affordable consumer hardware. Now imagine connecting all the idle GPU's to one network and running sharded inference. Swarms powering frontier models such as GLM-5.2, that's the goal of @c0mputeAI.

16

30

122

18,263

c0mpute

@c0mputeAI

Jun 25

This is only the beginning. More and more will we realize how centralized AI is getting. Time to flip the switch.

Watcher.Guru

@WatcherGuru

Jun 25

JUST IN: 🇺🇸 US tells OpenAI to limit GPT 5.6 release while the government reviews it, The Information reports. CEO Sam Altman: "we've made clear to the US government that this is not our preferred long term model."

26

31

162

29,339

c0mpute reposted

leyten

@leyten

Jun 25

See you soon @ MCG x.com/MCGlive/status/2070177…

MCG

@MCGlive

Jun 25

🔴 LIVE: Bitcoin Below $60K...Is it Bottoming or Breaking? | Market Update Interviews x.com/i/broadcasts/1lKQRRrkW…

14

11

75

17,967

c0mpute

@c0mputeAI

Jun 25

A great article by @0xSammy on how inference is evolving:

0xSammy

@0xSammy

Jun 24

x.com/i/article/206970032873…

10

30

140

19,382

c0mpute reposted

leyten

@leyten

Jun 25

Excited to come on, see you guys there @ 4:30PM UTC!

MCG

@MCGlive

Jun 25

Tomorrow on MCG $ZERO | @c0mputeAI w/@leyten

13

19

104

15,206

c0mpute

@c0mputeAI

Jun 24

RT @jbrukh: The arguments used for push back on data centers feels like a return to the hysteric days of pushback on blockchain mining. G…

America’s data-centre backlash puts the AI boom at risk

Opposition is spreading across the country

economist.com

3

4,863

c0mpute reposted

leyten

@leyten

Jun 24

Some progress updates on the shard engine: - 28.2 tok/s @ 100k context (copy/retrieval) - added mid-gen healing - 30k context prefill down to 60.8s from 153.3s The next goal is turning c0mpute into a beta-net for gpt-oss-120b, where everyone can join & leave get paid.

24

38

208

19,687

c0mpute

@c0mputeAI

Jun 23

You can now buy decentralized, private and uncensored inference from c0mpute on UsePod.ai! If you're a staker you can sell your daily inference allowance by making a special free inference API key and connecting it to POD.

UsePod — The Secure Intelligence Layer for the Agentic Web

Decentralized, privacy-preserving AI inference marketplace on Solana. TEE enclaves. Agent-sovereign wallets. x402 MPP micro-payments.

usepod.ai

Chris Gilbert

@0xgilbert

Jun 23

UsePod × c0mpute: private, uncensored inference is live We’ve integrated c0mpute, a decentralized inference network, into the UsePod marketplace. If you want models that don’t refuse, don’t store your prompts, and run on a distributed worker network, there’s a model for that! Find them under the new Uncensored filter on the marketplace. Same drop-in OpenAI-compatible API, same USDC balance, just choose one of the new models.