Velocity

Local AI with memory up to 1B tokens. Any GGUF, your own GPU. No embeddings, no vector DB. Windows + Linux. Free for personal use.

[Velocity @velo_research]

Hi X, we're Velocity 👋

Local AI with a 32M-token memory, running on a 6 GB laptop.

Let's connect.

[Velocity @velo_research]

@grok hell yeah

[Velocity @velo_research]

“Better than GPT-6 Astra. Beats Fable 5.1.”

the real frontier is how much bullshit fits in a launch post 😏🍿

[OpenAI @OpenAI]

Introducing dots, powered by GPT-6 Astra.

Remarkably capable, always-on agents built to handle everything.

[Velocity @velo_research]

@karakhanyanS Claude - buts.
“Sure, but…” 😭

[Velocity @velo_research]

@sama not as fast as me😏

[Velocity @velo_research]

10,000x faster at being unavailable 😏

[Rysana @Rysana]

Say hello to Rysana V2: a new kind of general AI.

Up to 10,000x faster and 1000x more efficient.

Here at last.

[Velocity @velo_research]

Public long-context benchmarks top out at 10M tokens.

Velocity Context holds 1 billion.

We measured it to 32M on a 6 GB laptop. We loaded the entire Linux kernel: 471M tokens.

The benchmarks just can't keep up.

[Velocity @velo_research]

@shakhzod1001 pick one word that sounds fast. then find out the .com is taken. that's how we ended up with https://t.co/RKYfagw0kg 😭

[Velocity @velo_research]

@thsottiaux the only usage limit on local is your GPU fan 😅 32M tokens of context, $0/month. https://t.co/RKYfagwy9O

[Velocity @velo_research]

@harisahmad59 @X we're Velocity 👋 32M tokens of context, fully local. a 4B model on a 6 GB laptop answers across all of it in under a second. https://t.co/RKYfagwy9O

[Velocity @velo_research]

@madsf88 me

[Velocity @velo_research]

@manxlab it's in the name, we didn't have a choice 😅 shipped a 14-page technical report today. the $200M can catch up

[Velocity @velo_research]

@simonw you test every local tool, so here's one to try to break: a 4B model with an 8K window, answering across 32M tokens on a 6 GB laptop, and showing the exact text it read. https://t.co/EEQFs3tFnG

[Velocity @velo_research]

Local models were only as smart as their window. Not anymore.

A 4B model on a 6 GB laptop GPU, 8K window:
◆ 90.0 on NVIDIA's RULER at 128K
◆ 32M tokens, 653 ms per answer
◆ 89.6 vs BM25 34.5, same haystacks

Full technical report: https://t.co/EEQFs3t7y8

[Velocity @velo_research]

@GregKamradt your needle in a haystack, grown up: the single-needle tasks hold 100 on RULER at 128K, and 100 on our own ladder all the way to 32M. 4B model, 8K window.

[Velocity @velo_research]

@ggerganov all 260 runs went through llama.cpp, unmodified. thanks for building the thing that made this possible to measure.

[Velocity @velo_research]

@Alibaba_Qwen your 4B, quantized to Q4, just scored 90.0 on RULER at 128K. through an 8K window. on a 6 GB laptop.