File size: 4,279 Bytes
cfb84ae
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
---
base_model: mistralai/Leanstral-2603
library_name: mlx
tags:
  - rotorquant
  - kv-cache-quantization
  - mlx
  - 8bit
  - weight-quantization
  - leanstral
  - lean4
  - formal-proofs
  - theorem-proving
  - quantized
  - apple-silicon
  - mistral
  - moe
license: apache-2.0
pipeline_tag: text-generation
---

# Leanstral-RotorQuant-MLX-8bit

**8-bit MLX weight-quantized [Leanstral-2603](https://huggingface.co/mistralai/Leanstral-2603) with [RotorQuant](https://github.com/scrya-com/rotorquant) KV-cache quantization for Lean 4 formal proof generation on Apple Silicon.**

Leanstral is the first open-source AI agent purpose-built for Lean 4 formal proofs -- generating both executable code and machine-checkable mathematical proofs. This variant combines **dual compression**: 8-bit MLX weight quantization for reduced model size plus RotorQuant KV-cache quantization for efficient long-context inference with faster prefill and decode.

Approximate model size: **~120 GB**

## Overview

This repository provides a **dual-compressed** configuration: MLX 8-bit weight quantization reduces the static memory footprint, while RotorQuant compresses the KV cache at runtime with superior throughput. Together, they enable running Leanstral on high-memory Apple Silicon machines.

| Spec | Value |
|------|-------|
| Base model | [mistralai/Leanstral-2603](https://huggingface.co/mistralai/Leanstral-2603) |
| Architecture | Mistral MoE (~119B parameters, 7 consolidated shards) |
| Weight quantization | 8-bit (MLX) |
| KV-cache quantization | RotorQuant |
| Weight memory | ~120 GB |
| Runtime | MLX (Apple Silicon) |
| License | Apache 2.0 |
| Use case | Lean 4 formal verification, theorem proving, mathematical proofs |

## Quickstart

```python
from mlx_lm import load, generate

model, tokenizer = load("majentik/Leanstral-RotorQuant-MLX-8bit")

prompt = "Prove that for all natural numbers n, n + 0 = n in Lean 4:"
response = generate(
    model,
    tokenizer,
    prompt=prompt,
    max_tokens=512,
)
print(response)
```

## What is RotorQuant?

[RotorQuant](https://github.com/scrya-com/rotorquant) is a rotation-based KV cache quantization method that applies learned Clifford algebra rotations before quantizing the key-value cache. Key results:

- **5.3x faster prefill** compared to TurboQuant baseline
- **28% faster decode** throughput
- **Perplexity: 6.91** vs 7.07 for TurboQuant (lower is better)

Combined with MLX 8-bit weight quantization, this dual compression approach makes it feasible to run Leanstral's ~119B parameter model on Apple Silicon hardware with excellent throughput.

## KV-Cache Quantization Comparison

| Method | Prefill Speed | Decode Speed | Memory Savings | Reference |
|---|---|---|---|---|
| **TurboQuant** | Baseline | Baseline | High | [arXiv: 2504.19874](https://arxiv.org/abs/2504.19874) |
| **RotorQuant** | 5.3x faster | 28% faster | High | [GitHub](https://github.com/scrya-com/rotorquant) |

## Memory Estimates

| Component | Estimate |
|-----------|----------|
| Model weights (8-bit) | ~120 GB |
| KV-cache | Reduced via RotorQuant |
| Recommended hardware | Mac Studio M2/M3/M4 Ultra (192 GB+) or Mac Pro |

## Lean 4 Use Case

Leanstral excels at:
- **Formal verification** -- generating machine-checkable proofs of mathematical theorems
- **Theorem proving** -- interactive and automated proof search in Lean 4
- **Code generation** -- writing verified Lean 4 programs with correctness guarantees
- **Proof repair** -- fixing incomplete or broken proof scripts

## See Also

- [mistralai/Leanstral-2603](https://huggingface.co/mistralai/Leanstral-2603) -- Base model
- [majentik/Leanstral-RotorQuant](https://huggingface.co/majentik/Leanstral-RotorQuant) -- Full-precision weights + RotorQuant KV cache
- [majentik/Leanstral-RotorQuant-MLX-4bit](https://huggingface.co/majentik/Leanstral-RotorQuant-MLX-4bit) -- MLX 4-bit + RotorQuant
- [majentik/Leanstral-RotorQuant-MLX-2bit](https://huggingface.co/majentik/Leanstral-RotorQuant-MLX-2bit) -- MLX 2-bit + RotorQuant
- [majentik/Leanstral-TurboQuant-MLX-8bit](https://huggingface.co/majentik/Leanstral-TurboQuant-MLX-8bit) -- MLX 8-bit + TurboQuant
- [RotorQuant GitHub](https://github.com/scrya-com/rotorquant)
- [MLX Framework](https://github.com/ml-explore/mlx)