Leanstral-TurboQuant-MLX-8bit

8-bit MLX weight-quantized Leanstral-2603 with TurboQuant 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 TurboQuant KV-cache quantization for efficient long-context inference.

Approximate model size: ~120 GB

Overview

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

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

Quickstart

from mlx_lm import load, generate

model, tokenizer = load("majentik/Leanstral-TurboQuant-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 TurboQuant?

TurboQuant (arXiv: 2504.19874) is a KV-cache quantization method that compresses the key-value cache used during autoregressive generation. By quantizing the KV cache to lower precision, TurboQuant reduces memory consumption proportionally to context length. 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.

KV-Cache Quantization Comparison

Method	Prefill Speed	Decode Speed	Memory Savings	Reference
TurboQuant	Baseline	Baseline	High	arXiv: 2504.19874
RotorQuant	5.3x faster	28% faster	High	GitHub

Memory Estimates

Component	Estimate
Model weights (8-bit)	~120 GB
KV-cache	Reduced via TurboQuant
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 -- Base model
• majentik/Leanstral-TurboQuant -- Full-precision weights + TurboQuant KV cache
• majentik/Leanstral-TurboQuant-MLX-4bit -- MLX 4-bit + TurboQuant
• majentik/Leanstral-TurboQuant-MLX-2bit -- MLX 2-bit + TurboQuant
• majentik/Leanstral-RotorQuant-MLX-8bit -- MLX 8-bit + RotorQuant (faster prefill/decode)
• TurboQuant paper
• MLX Framework

Quant trade-off (MLX lane)

Bits	Approx size	Use case	Recommendation
2-bit	~31 GB	Aggressive quantization	Very low-RAM Macs
3-bit	~43 GB	Lossy but small	Low-RAM Macs
4-bit	~50 GB	Balanced default	Recommended for most Macs
5-bit	~60 GB	Higher fidelity	Quality-sensitive
6-bit	~71 GB	Approaching FP16 quality	High-fidelity
8-bit	~90 GB	Near-lossless reference	Fidelity-critical work

(Current variant — 8bit — is bolded.)

Variants in this family

(Showing 8 sibling variants under majentik/leanstral-*. The current variant — TurboQuant-MLX-8bit — is bolded.)

Variant	Runtime	Approx size	Use case
RotorQuant	runtime modifier	n/a	KV-cache root (weight-agnostic)
RotorQuant-MLX-2bit	mlx-lm	card-only	Apple Silicon, smallest
RotorQuant-MLX-4bit	mlx-lm	card-only	Apple Silicon balanced
RotorQuant-MLX-8bit	mlx-lm	card-only	Apple Silicon reference
TurboQuant	runtime modifier	n/a	KV-cache root (weight-agnostic)
TurboQuant-MLX-2bit	mlx-lm	card-only	Apple Silicon, smallest
TurboQuant-MLX-4bit	mlx-lm	card-only	Apple Silicon balanced
TurboQuant-MLX-8bit	mlx-lm	card-only	Apple Silicon reference