Then we promptly committed these changes as open source and contributed back to the community. Now anyone with B200s can run this. Admittedly theyre not that easy to get. https://github.com/comput3ai/c3-vllm

Don't have any B200s or dont feel like running them? That's okay too. We have some.

The token is $COM. The CA is J3NrhzUeKBSA3tJQjNq77zqpWJNz3FS9TrX7H7SLKcom

Some projects are just built differently. 20 min turnaround on this feature request. https://github.com/comput3ai/c3-docker-images/commit/5989a88a2086fc432845cdb7dd8c465ee63f338c

Qwen 3 Coder 30B is already very impressive. Expectations are high for the 480B. Try one right now https://launch.comput3.ai

Join us if you're around this morning.

🇮🇳 Aya AI Hackathon Bangalore 🇮🇳

Going live now! 🎥

Aya AI Hackathon x Comput3 Livestream – dive into MCP builds, free GPU hours & more.

🔗 https://youtube.com/live/7CbfBiQvhv0?feature=share

Bring your questions, grab some tips, and level up your #AyaAIHackathon project.

Decided to share this with everyone already so no one is surprised when we start posting about it. When we launched COMPUT3 we pronounced it "compute" the idea was to fair launch on @autodotfun and give a subtle hint at what we're building. We thought there should be a web3 native compute platform thats build completely differently and from the ground up.

Thanks to your incredible support, our platform, capabilities, infrastructure, and especially our community, have grown by leaps and bounds, far beyond our wildest dreams. We didn't know what to expect, but we're truly humbled and grateful for where we are right now! None of this would be possible without you.

We posted this earlier this week, and, sadly, it's not a joke. This really happened: https://x.com/comput3ai/status/1952240139390066721?t=Is16AhDZsog1sZX_aMNagg&s=19

The feedback we get from non-web3 companies is that our branding doesn't appeal to them. We're a token that's all about community, and we do this for you guys at least as much as we do it for ourselves. We are working to build a platform where AI agents can themselves book compute. That has to be in crypto. But these non-Web3 folks just want to use our tech; they don't care how cyberpunk our website looks. In fact, it's a reason for them to look away.

All of that is to say, we're announcing today that we're launching compute3.ai (with an 'e') as the corporate face of comput3.ai the token. We'll use a unified tech stack, and we'll make sure anything we build for web2 is available for web3. You can also follow us @compute3ai - we appreciate every one of you who joins us there.

We know our roots. We're 🌋 all the way. But to facilitate revenue outside of Web3, this is a necessary step in our growth, and we're so thankful for your understanding and enthusiasm as we evolve. It feels like we've grown up, together. And we're just getting started.

All the Web3 content, live streams, and fun will stay here, but things you can buy and pay for in fiat will be over at @compute3ai. Thank you for being part of this journey!

We just updated ollama to be able to run gpt-oss. Crazy to see even OpenAI get shamed into creating open weight models. The crazy part is arguably, OpenAI's most widely used model is Whisper. Which is their open model for audio transcription.

Open source is the way. Follow the infra 🌋

OpenAI is pitching gpt-oss as their edge model. What's edge? It means it can run locally on your machine not on their cloud. This means more things can run these models. They're more compact and optimized.

But people forget models have to be trained and what are they trained on? B200s and H200s. Training is always most efficient on the biggest hardware available. Always.

https://developer.nvidia.com/blog/delivering-1-5-m-tps-inference-on-nvidia-gb200-nvl72-nvidia-accelerates-openai-gpt-oss-models-from-cloud-to-edge/

Right now most distributed and decentralized training networks on Web3 are falling back to single consumer 3090s and 4090s. This makes sense while they're on testnet. Side note, Psyche runs on H100s on testnet, which puts them in a different league. But what happens when all these networks go to main net? What happens when they want to train real models? Where will they get B200s, H200s, H100s?


WHERE ARE THE GPUS, LEBOWSKI?!

Only on $COM 🌋

CA is J3NrhzUeKBSA3tJQjNq77zqpWJNz3FS9TrX7H7SLKcom

The world's fastest gpus for AI training.
Large single gpus for large models with large contexts.
Models you can run locally on your data.
ARE YOU GETTING IT?

https://openai.com/index/gpt-oss-model-card/

🌋 $COM CA: J3NrhzUeKBSA3tJQjNq77zqpWJNz3FS9TrX7H7SLKcom

We had to enlist Steve to explain this to you. ARE YOU GETTING IT?

This was built using open source models and rendered on our gpu network. Try it out right now through our bot 👉 https://t.iss.one/C3PortraitBot

Support us in what we're buidling and get GPU hours for workflows like this every month by staking $COM.

🌋 $COM CA: J3NrhzUeKBSA3tJQjNq77zqpWJNz3FS9TrX7H7SLKcom

One of the uses of AI where we expect the most explosive growth. Code security automation ... where you probably don't want a closed source model thats using your code for training data. Guess who's building just that, private open source models with claude code compatibilty? 🌋🌋🌋

The best and most powerful open source models are coming to @comput3ai. Stake $COM to get access to Kimi K2 and Qwen 3 Coder, as well as GPU hours every month. Or just help us dominate this growing sector.

The CA is: J3NrhzUeKBSA3tJQjNq77zqpWJNz3FS9TrX7H7SLKcom

Quick PSA. We were made aware by our community that the additional streamflow $COM rewards weren't being distributed correctly to stakers. This is in addition to the GPU hours. We've decided to remedy this by airdropping another 5M $COM to those who staked on streamflow, but weren't able to get rewards! We will launch a second streamflow contract without rewards as these were intended to get more people to stake initially while we were getting everything else up and running.

If you're already staking, expect to get airdropped your rewards. This should happen starting next week and until 8/31/2025. You don't need to restake to receive GPU rewards. But be aware there will be no more $COM rewards distributed via the old streamflow contract. The new streamflow contract will remove $COM rewards completetly to prevent confusion among stakers.

Thank you 🌋

$COM streamflow staking rewards are live. Thanks for bearing with us while we got the airdrop setup. https://app.streamflow.finance/airdrops/solana/mainnet/HjCjVANUof9MxdxFVSkQKaMZtDstGbTvp1mFAiisJNyR (updated URL)

Now go explain to your mom what a B200 is. $COM fam🌋

A few folks reached out privately asking for a breakdown of this chart, so let's dive in and unpack it properly.

First off, the knee-jerk reaction might be: "Oh, this is just a post-COVID shift." Not quite. COVID's major impacts wrapped up around 2022, while ChatGPT and the broader AI boom really hit the headlines in 2023. So, timing-wise, that doesn't add up.

What we're looking at here isn't total servers installed or even the number of data centers built—it's total spending, in dollars. The Y-axis tells the story: raw capital outlay.

Now, you might wonder, "Servers and PCs cost about the same today as they did in 2022—where's all this extra money going?" The key insight? These aren't your standard data centers or everyday servers. We're talking AI data centers packed with specialized AI servers—like NVIDIA's B200, H200, and H100 GPUs. These beasts are exponentially more expensive than traditional hardware, which is exactly what creates that sharp "kink" in the chart.

Zoom in further, and you'll notice the monthly growth is remarkably linear. If you're familiar with manufacturing, this makes perfect sense: it mirrors TSMC's steady ramp-up in production capacity to churn out GPUs for NVIDIA. Supply chains don't scale overnight—they build predictably, and that's reflected here.

This chart essentially captures our project's core thesis in visual form. AI doesn't have to break the bank. There's already an oversupply of GPUs out there, and everything in AI—from training to inference—can be handled efficiently in a distributed way. We're all about running open-source models, sourcing hardware globally, and delivering massive bang for your buck. But it's bigger than cost savings: it's about true control, uncensored access, and ownership over your AI workflows. We build everything in the open and aim to be the premier platform for hosting, fine-tuning, and deploying models.

And what about the top line? Sure, you could frame it as "AI replacing humans." But we see the brighter side: AI supercharges productivity, making people 10x more efficient—without chaining them to an office desk. It's empowerment, not displacement.

What do you all think? Let us know in the comments.

Today, we squashed a couple of bugs in race conditions that only occur at our scale. We also shipped a couple of improvements to our custom container orchestration layer. This will improve security, isolation and uptime and will work better on desktop PCs and not just 8xB200's.

That's right, we always make sure our stack runs on desktop PCs and even on Windows. You'll know why soon enough.

🏆 Aya AI Hackathon – 🇮🇳 Bangalore Winners Announced 🏆

Our Bangalore winner tweet is live on X!

🔗 Tweet: https://x.com/theayawallet/status/1956626774806479207

How to help (takes 30–60s):
1️⃣ Like ❤️ + RT 🔁 the tweet
2️⃣ Quote-tweet with a shout-out to your favorite team
3️⃣ Tag partners & IIIT-B to boost reach
4️⃣ Drop a congrats comment 🙌

Let’s get these builders the spotlight they deserve! 📣

Another awesome event by Aya!

These teams built their projects on @comput3ai and we gave them some GPU hours so they can keep cooking.

1️⃣ Team Hackxbot — Multi-Chain AI-Powered Crypto Manager
2️⃣ Team Delta — SnackGpt
3️⃣ Suyash — AI-Powered Crypto Assistant

Today's the big day, we just deployed kimi-k2 and qwen3-coder:480b to our api. If you want to play around with this, hop on our discord: https://discord.com/invite/DmaHAdGcNy

Alright @everyone it's the big day.

This is work in progress. We are doing a ton of stuff that you can't see, but it is up and we would appreciate testing and feedback. The inferencing API is now running on B200's. Right now everyone has access to all the models, we'll be closing this down as we go along.

We'll be launching paid subscriptions, we had a lot of discussions and honestly these models are so powerful that we'll probably be targeting $99/mo, but we'll start with $75/mo as a welcome offer. You won't have to use phantom, it'll be email and stripe (we'll do crypto too don't worry). We will be giving stakers access to what we consider reasonable limits (input/output tokens per month), but they'll be less than the subscriptions at least initially. Eventually we hope we can just map stakers to these plans, but for the moment we just added a bunch of 24/7 infra, which we're running in parallel to the GPU workloads most of you are familiar with.

The models on the api include the hermes3:70b and llama3:70b, both now running at 64K context (thanks B200s), but the big news is kimi-k2 and qwen3-coder. Right now everyone with an api key has access to these models. We'll be limiting them to stakers and subscribers later this week.

Keep in mind these are monstrous models and you might hit the Cloudflare's request timeouts after 100 seconds. That's not us - that's a cloudflare "feature". We're working on the api.compute3.ai (with an 'e') deployment as we speak, and that will be naked (no cloudflare, just a couple of load balancers) so it won't have that issue. We'll be adding instructions on how to set all of this up with claude code, qwen code and probably codex as well. In the mean time feel free to test, api.comput3.ai is now running a litellm-proxy so it should handle most of the shenanigans that the coding agents need. However, we're not routing /anthropic/* (or /openai/ for that matter) to our models yet, we're looking for a reasonable way to be able to leave you control over which goes where instead of picking one for you. I do believe overloading the /model should work, but I haven't tested this yet.

Enjoy and let us know how it goes. Here's some curls to get you started.

$ curl -X GET https://api.comput3.ai/v1/models | jq
{
  "data": [
    {
      "id": "hermes3:70b",
      "object": "model",
      "created": 1677610602,
      "owned_by": "openai"
    },
    {
      "id": "kimi-k2",
      "object": "model",
      "created": 1677610602,
      "owned_by": "openai"
    },
    {
      "id": "qwen3-coder:480b",
      "object": "model",
      "created": 1677610602,
      "owned_by": "openai"
    },
    {
      "id": "llama3:70b",
      "object": "model",
      "created": 1677610602,
      "owned_by": "openai"
    }
  ],
  "object": "list"
}

You can access these as a normal /v1 API.

curl -X POST https://api.comput3.ai/v1/chat/completions \
-H "Authorization: Bearer c3_api_YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
      "model": "kimi-k2",
      "messages": [
        {
          "role": "system",
          "content": "You are a helpful assistant."
        },
        {
          "role": "user",
          "content": "What is 2+2?"
        }
      ]
    }'
{"id":"chatcmpl-007e9d34e4474e78902934247b1f72bc","created":1755501320,"model":"kimi-k2","object":"chat.completion","system_fingerprint":null,"choices":[{"finish_reason":"stop","index":0,"message":{"content":"2 + 2 = 4","role":"assistant","tool_calls":null,"function_call":null},"provider_specific_fields":{"stop_reason":163586}}],"usage":{"completion_tokens":8,"prompt_tokens":24,"total_tokens":32,"completion_tokens_details":null,"prompt_tokens_details":null},"service_tier":null,"prompt_logprobs":null,"kv_transfer_params":null}

vibe is live 🌋🌋🌋

export ANTHROPIC_BASE_URL=api.comput3.ai
export ANTHROPIC_AUTH_TOKEN=c3_api_YOUR_C3_API_KEY
export ANTHROPIC_MODEL=kimi-k2