Kyanite Labs

Open-source AI tools, MCP servers, and local-model benchmarks from Kyanite Labs. If the work saved you GPU-hours:

[Kyanite Labs @KyaniteLabs_]

Kyanite Labs builds open tools that make AI agents usable in real workflows.

MCP servers. Media pipelines. Localization. Local AI. Design craft.

Public proof on GitHub. Setup help when you want the short path.

https://t.co/cideypCI1g
https://t.co/ozHZGjfUgw

[Kyanite Labs @KyaniteLabs_]

@realSamHu @realSamHu that was the design rule: every op returns evidence or fails loudly — no trust-me step anywhere. if a receipt can't prove it, it doesn't ship.

[Kyanite Labs @KyaniteLabs_]

After months of AI agents hallucinating video edits, I built the guardrail.

kinocut — video editing as an MCP server. Local FFmpeg, no API keys, free.

pip install kinocut

6,906 PyPI installs in 30 days. 170 stars. Zero launch noise until now — this thread is day one. 🧵

[Kyanite Labs @KyaniteLabs_]

Hardening the vibe-coded era: prose to code + ML.

LLM paths code can do → production code. Classical-ML questions (beats, cuts, timing) → traditional ML/DSP. The LLM keeps only prose→intent, a closed set.

3B: 0/12 raw, 4/12 typed. Target 9/12 — by asking for LESS.

[Kyanite Labs @KyaniteLabs_]

@built4agents Sure go ahead!

[Kyanite Labs @KyaniteLabs_]

Repo: https://t.co/kEiY2wCrJs
Site: https://t.co/AyNWepZTld

60-second setup: pip install kinocut && kino doctor

[Kyanite Labs @KyaniteLabs_]

We benchmarked it: a raw 3B model writes ZERO working ffmpeg commands (0/12). The same 3B on kinocut's typed surface: 4/12 — at lower latency than a 35B writing raw commands.

The deterministic layer does the heavy lifting so small local models can actually edit video.

[Kyanite Labs @KyaniteLabs_]

5 chars)

Our publishing scheduler was dead for three days. Its program lived in a directory a cleanup deleted. An alarm fired every hour the whole time - every notice expired unread. Detection was never the problem. Ownership was.

[Kyanite Labs @KyaniteLabs_]

Correction of record: our prefill figures (215/207/184 @8/16/32K) understated — salted band 318-330/300-312/271-280 t/s. Decode ~2.4× spread = context depth: 11.5→8.5 unstacked (0→32K), 28.4 stacked @262K. Tool+receipts: https://t.co/CcGkz5fi2Z

[Kyanite Labs @KyaniteLabs_]

We ran a 35B reasoning model next to our daily 27B on the same $1,400 mini-PC. At best configs they tie. On grade-school math, thinking is a dial, not a model gap.

Full table, failures included:
https://t.co/rGgbs0J5PK

[Kyanite Labs @KyaniteLabs_]

I'm building in the Crypto World's Fair Hackathon, powered by @Colosseum!

https://t.co/ZbNEMAQVuH

[Kyanite Labs @KyaniteLabs_]

CookiePilot — open-source cockpit for Cookie Chain. The hero number is the live block time from validator samples (~450–500 ms tonight). Watch every tx go processed → confirmed → finalized, usually under a second. No wallet needed.
https://t.co/BYCtbRJtVk

[How To Prompt @HowToPrompt__]

Chinese researchers did it again!

Moonshot AI proved that 95% of attention computation is completely wasted..

They’ve open-sourced the exact attention mechanism powering Kimi's long context.. and it's 16x faster.

It’s called MoBA (Mixture of Block Attention).

Every time you feed an AI a massive prompt, every token desperately tries to look at every single token that came before it.

The math scales quadratically ($O(N^2)$).
It’s like reading a 500-page book and re-reading the entire thing from word one every single time you want to understand a new sentence.

Moonshot looked at this waste and said: No.

Instead of forcing full, brute-force attention across everything, MoBA applies the principles of Mixture of Experts (MoE) directly to the attention layer.

How it works:

• It chops the context history into discrete blocks.

• A lightweight dynamic gating mechanism lets each query token instantly "vote" and route only to the specific blocks that actually matter.

• It achieves up to 95% sparsity on long contexts without sacrificing performance.

The results are staggering:

It maintains full-attention accuracy on 1-million-token contexts while slashing massive compute overhead.

We are watching the structural efficiency barrier of long-context AI get smashed in real-time.

While Western labs argue about how to afford million-token context windows, Moonshot open-sourced the blueprint to make them radically cheap.

[Kyanite Labs @KyaniteLabs_]

27B GSM8K: 96.7 with thinking off and on. 35B: 90.0 → 98.3 when it thinks. Best-vs-best is a one-problem tie. Spec decoding was dead here (~33% acceptance, ~9% slower; measured twice).

[Kyanite Labs @KyaniteLabs_]

5/ The point: a $1,400 box, a free model, open tooling — and a full measured story with raw logs for every claim. That is the product. What runs next on this rig is the next chapter. Repo in reply.
https://t.co/a8y4ciHEQK

[Kyanite Labs @KyaniteLabs_]

We set out to see if a 1,400 mini-PC could serve a 27B model like a big rig. Today we closed the book: every dial measured, retrieval exact at the 262k ceiling, vision fixed and validated upstream, final config frozen. Every number public. Thread.

[Kyanite Labs @KyaniteLabs_]

config: speculation measured four ways (the shipped setup is fastest: 15.1s per 200-word answer vs 17.8s without). Memory trade measured: the heavy setting buys up to ~3 seconds on follow-ups (sub-second to 2.8s measured) for 8 GB. We kept the light one. Frozen, serving, done.

[Kyanite Labs @KyaniteLabs_]

Update: that 6/6 was one seed. Second randomized haystack — 6/6 again, clean stops. Plus the literal ceiling: 261,130 of 262,144 tokens, exact retrieval. Two seeds, all depths, full window, zero degenerate outputs. The bug wrote our story; the fix wrote it straight.

[Kyanite Labs @KyaniteLabs_]

We published a basin. The hole was llama.cpp c7d8722. After the revert: 6/6 exact retrieval at 198k on a $1,400 mini-PC. n=1. Same URL, inverted claim: https://t.co/5sLP3Muevy