File size: 3,281 Bytes
a0e777e | 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 | ---
base_model: mistralai/Mistral-Small-4-119B-2603
library_name: transformers
license: apache-2.0
tags:
- rotorquant
- kv-cache-quantization
- mistral
- moe
- sparse-moe
- multimodal
- quantized
- 256k-context
- thinking
pipeline_tag: text-generation
---
# Mistral-Small-4-119B-RotorQuant
**KV cache quantization for Mistral Small 4 using RotorQuant** -- 5.3x faster prefill, 28% faster decode, with near-lossless quality (perplexity 6.91 vs 7.07 baseline).
This repository provides RotorQuant KV cache quantization support for [mistralai/Mistral-Small-4-119B-2603](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603). Model weights are unchanged (FP16); only the KV cache is quantized during inference.
## Model Specs
| Property | Value |
|---|---|
| Base Model | Mistral Small 4 (March 2026) |
| Total Parameters | 119B |
| Active Parameters | 6.5B per token (Sparse MoE) |
| Architecture | Sparse MoE -- 128 experts, 4 active per token |
| Context Length | 256K tokens |
| Modality | Text + Images (multimodal) |
| Capabilities | Thinking / reasoning, tool use, multilingual |
| License | Apache 2.0 |
| Quantization | KV cache only (RotorQuant) |
## What is RotorQuant?
[RotorQuant](https://github.com/scrya-com/rotorquant) is a rotation-based KV cache quantization method that applies learned rotations before quantizing the key-value cache. Key results:
- **5.3x faster prefill** compared to unquantized baseline
- **28% faster decode** throughput
- **Perplexity: 6.91** vs 7.07 for unquantized (lower is better -- RotorQuant actually improves quality due to outlier suppression)
- Default 3-bit quantization with minimal quality loss
## Memory Estimates
| Component | FP16 Baseline | RotorQuant 3-bit |
|---|---|---|
| Model Weights | ~238 GB | ~238 GB |
| KV Cache (256K ctx) | ~32 GB | ~6.5 GB |
| **Total** | **~270 GB** | **~244.5 GB** |
> **Note:** This is a Sparse MoE model -- only 6.5B parameters are active per token, so inference is fast despite the 119B total parameter count.
## Quickstart
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from turboquant import IsoQuantCache
model_id = "majentik/Mistral-Small-4-119B-RotorQuant"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype="auto",
device_map="auto",
)
# Enable RotorQuant KV cache
cache = IsoQuantCache(model)
messages = [
{"role": "user", "content": "Explain sparse mixture-of-experts architectures."}
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)
outputs = model.generate(
inputs,
max_new_tokens=512,
past_key_values=cache,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```
## See Also
- [mistralai/Mistral-Small-4-119B-2603](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603) -- Base model
- [majentik/Mistral-Small-4-119B-TurboQuant](https://huggingface.co/majentik/Mistral-Small-4-119B-TurboQuant) -- TurboQuant KV cache variant
- [majentik/Mistral-Small-4-119B-RotorQuant-MLX-4bit](https://huggingface.co/majentik/Mistral-Small-4-119B-RotorQuant-MLX-4bit) -- MLX 4-bit weight-quantized + RotorQuant
- [RotorQuant GitHub](https://github.com/scrya-com/rotorquant)
|