Decompute AI

Decompute is an artificial intelligence (AI) startup focused on decentralizing AI development and infrastructure.

[Decompute AI @DecomputeAI]

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: https://t.co/5e2NgTnk0d
The full power of AI is now in your hands.

[Hina Dixit @hinadixit]

93.75% KV reduction, with our unique memory model architecture Kiran, with just additional 100 MBs, no pre-training or fine-tuning required. We scored an almost flat accuracy of 84% at 10M tokens! Costed me only $39! Can you imagine how we can shift the GPU memory shortage?

[Hina Dixit @hinadixit]

That’s why I build Kiran, a perfect flat score of 84% at 10M or 1M, with linear attention! Phew! Let model manage its own memory with just additional 100 MBs, and answer question from 10 years, like it was 10 mins ago. Love to see, people feel this problem intensely.

[Hina Dixit @hinadixit]

My pov is simple. We need more compute, but at an affordable price, where we don’t make AI a bubble. Excited to introduce our heterogeneous inference’s V0.

[Hina Dixit @hinadixit]

Introducing Kiran-10M: 84.0% average accuracy on BABILong qa1–qa5 at 10M tokens — and 85.6% at 0k. Same frozen 4B-class open-weights LLM, same memory, every length in between inside 84.0–86.4. Ours is the flat one. 1/7

[Hina Dixit @hinadixit]

Oh, and one more thing. It costed me only $39.

[Hina Dixit @hinadixit]

Thank you for reviewing BF1, James! Checkout our results for Qwen3.8-27B at 1M context as well.

https://t.co/iZk2KtMAJc

[Hina Dixit @hinadixit]

Thank you for covering BF1, Chad! Appreciate it! 🫡

https://t.co/uA2IdYOdzZ

[Hina Dixit @hinadixit]

BF1, training-free for Qwen3.8-27B at 1M context. Mind == Blown. 🤯

[Hina Dixit @hinadixit]

Bigger ideas, lesser compute. No pre-training required. Strict O(n log n).

[Hina Dixit @hinadixit]

Some early numbers from our new attention method Manojava, O(n) strictly linear on an adversarial dataset. This is completely algorithmic and different than our BF1 implementation. Both methods do not require pre-training.

[Hina Dixit @hinadixit]

For a quick peek. 🎁

[Hina Dixit @hinadixit]

It’s overwhelming to see how many people messaged me to appreciate BF1, and how it does not need model to be pre-trained to reduce attention to O(n log n). I truly enjoyed each conversation and felt inspired. Thank you! 🙏🏻

[Hina Dixit @hinadixit]

The distinction is simple: BF1 does not need to search the entire history to decide where to attend. No learned router. No history-wide similarity search. No retrieval system bolted onto the model. No need to pretrain a new architecture from scratch.

[Hina Dixit @hinadixit]

BF1 removes that search altogether.
We didn’t make dense attention faster. We stopped doing most of it.

Paper: https://t.co/L5yu4t14n3

[Hina Dixit @hinadixit]

DeepSeek DSA takes a different approach: its Lightning Indexer cheaply scores each query against preceding tokens, then runs the expensive attention computation only over the selected top-k. That makes the core attention sparse—but the indexer itself remains O(n²) in prefill.

[Hina Dixit @hinadixit]

Today I’m introducing BF1, a new causal multiscale attention operator for long-context Transformers. BF1 replaces quadratic all-pairs attention with deterministic O(n log n) communication At 32K: 10.91× faster attention · 26.98× fewer interactions · 15.3% lower whole-model TTFT.

[Hina Dixit @hinadixit]

Introducing my latest work: KIRAN gives a model writable state it can revise mid-reasoning. Same weights, workspace off→on: 60.3% → 79.1%.

The next frontier is state. Not more memory - memory that can change the model’s mind.

[Hina Dixit @hinadixit]

Get your local model to learn from your Fable5 interactions, just a few clicks away. check out https://t.co/jXQzkhB6Vn to set up a local model as student and Claude as the teacher. Distill your model, with simple steps!!

[Decompute AI @DecomputeAI]

@Thom_Wolf @cerebras 👏🏻👏🏻👏🏻

[Hina Dixit @hinadixit]

Download Decompute Claude Gateway, to save on tokens, and also enable your personal model to learn from your interactions/serve you offline.
https://t.co/jXQzkhB6Vn 🚀