Oskar Kocol

Founder @

[the tiny corp @__tinygrad__]

@RepCasar The people closes to AI technology would like others to slow down so they can maintain their lead. It's shameful that they would lie to try to get the government to help them with their business like this.

[Oskar Kocol @osquarski]

@grok is it my lying eyes or did just @baseten pull kimi k3 from @OpenRouter ? Why would that be?

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[Oskar Kocol @osquarski]

@codermatt @useblacksmith Oh I seee! So when the scheduler is down, the runners are also down ☠️

[Oskar Kocol @osquarski]

@lgrammel What's the best way to prove yourself?

[Oskar Kocol @osquarski]

@fekuuuu @ionet Macie program dla startupów w @ionet?

[Oskar Kocol @osquarski]

Finally a serious competitor to llama cpp and vllm. Let's go @basecmpt!

[Base Compute @basecmpt]

We’ve built the fastest inference runtime for LLMs on Apple Silicon, faster than MLX and llama.cpp.

BaseRT is up to 35% faster than Apple's MLX on token generation (decode) for small models, and 78% faster on prefill for bigger models.

Details below.

[Oskar Kocol @osquarski]

Tokenminning constitution is WIP - contribute on Discord

https://t.co/mJBIt2QKhk

[Oskar Kocol @osquarski]

You do not achieve token efficiency by asking engineers to write shorter prompts.

You achieve it by building infrastructure where token bloat physically cannot deploy

https://t.co/Nmb9Wnxs0S

[Oskar Kocol @osquarski]

5. GPTCache - Semantic cache for LLM apps

https://t.co/FSVm7GDDxO

[Oskar Kocol @osquarski]

1/5 tokenminning projects to star on Github:

1. PonyTail - You know him. Long ponytail. Oval glasses.
~54% less code (up to 94%) · ~20% cheaper · ~27% faster · 100% safe
https://t.co/H45DqUE7XU

[Oskar Kocol @osquarski]

4. Ollama - one-command local model runner

https://t.co/o1sqW63uIK

[the tiny corp @__tinygrad__]

@Polymarket Anthropic investors have two choices basically. Kick the fearmongerers out of the company, or start a new company without them. Scoop all the talent, wouldn't even take long. If you are a private company who talks like this, of course the government should crack down.

[Oskar Kocol @osquarski]

Dagster is the tool data teams use to schedule and run their data pipelines.

Think of it like a traffic controller for your data. It decides when to pull data in, when to transform it, and in what order.

We run ours 24/7 and it was quietly costing us a fortune. Until we found a fix that took about 10 minutes to implement. Full details;

https://t.co/0GNB5ozGPK

[Oskar Kocol @osquarski]

🚨 Dagster just raised prices 10x and most teams have NO idea how to fix it.

We cut our bill by 60% with one code change. No migration, no new tools, no self-hosting.

The culprit? A single line of Python that silently multiplies your credits 100x.

Here’s the fix 🧵

[Oskar Kocol @osquarski]

For those looking to lower their Dagster Cloud bill

https://t.co/PFCqJ9qWLv

[Oskar Kocol @osquarski]

Short recap of the FinOpsX 2026! Tokenomics is the FinOps word of 2026!

https://t.co/Wf5YQwGPli