Mostlysane AI

Run cutting-edge local AI on everyday hardware.
Zero monthly fees. Complete privacy. Your data stays local.
Create and process documents, images, and files entirely on your device.

Get Started Read the Research See Demos
3Research Tracks
Combined
512KContext Window
on 8GB GPUs
12/12CodeNeedle
Perfect Score
1.58Bit Ternary
Quantization

What We Built

Mostlysane is a curated stack of three research forks merged into a single llama.cpp build. The goal: make the latest model architectures and context sizes accessible on the hardware people actually own.

🧮

Ternary Quantization

Ported from a 1.58-bit ternary research fork. Weights become {-1, 0, +1} — 2 bits per parameter. Enables a new branch of models such as Bonsai 8B to run with minimal memory overhead.

💾

TurboQuant KV Cache

K-side compression via TurboQuant: turbo4 (4-bit), turbo6 (6-bit), mixed precision. V-side via turbo3 (Walsh-Hadamard + PolarQuant, ~3.125 bit). Reduces KV cache to 3Gb even at 512K context.

📊

Entropy-Adaptive Precision

Per-layer entropy profiling identifies which layers need full 8-bit K and which can use compressed types. Mixed precision keeps quality where it counts, memory savings everywhere else. Profiles for 5 models included, and a tool to create your own.

Why It Matters

New MoE architectures (Qwen3.6 35B A3B) unlock massive capability. But without custom load configurations and optimizations to caching, the KV cache consumes gigabytes at longer contexts — pushing them beyond reach of 8–12 GB GPUs. With our stack, MoE models put the important layers into GPU, reduce KV Cache without loss of quality and provide great performance, even with a massive 512k context window.

Before

Full-precision K/V cache at 512K context: ~11 GB
Qwen3.6 35B requires 24 GB+ GPU or reduced context

After

Calibrated entopy with TurboQuant K+V at 512K context: ~3 MB
Qwen3.6 35B fits on 8 GB GPU with full 512K context

Calibrated Entropy

Some things are more important that others. Mixed precision preserves quality on attention-critical layers.
Calibrated per model — set and forget.

Quality & Performance

We test configs against real-world multi-length recall & coding tasks and visual demos. Compression doesn't matter if quality suffers.

ConfigK BitsV BitsQualityPrefillGen512K Cache
Default F16 K & V, No optimisation 1616 Baseline 3600 t/s12 t/s ~10,500 MB
Q8_0 + Turbo3_0 with Entropy 83.125 Production 3372 t/s107 t/s ~3,000 MB
TurboQuant 6_0 + TurboQuant 3_0 63.125 Lossy 3168 t/s99 t/s ~3,200 MB
TurboQuant 4_0 + TurboQuant 3_0 43.125 Fragile 3268 t/s104 t/s ~2,700 MB

Benchmarked on AMD Rzyen 3800x, RTX 3070 Ti 8GB • Qwen3.6 35B A3B See visual comparisons →

Ready to Try It?

Generate a complete config tailored to your hardware in under 60 seconds.

Get My Config

Or dive straight in: curl -sSL mostlysane.ai/install.sh | bash