---
license: apache-2.0
base_model:
  - Qwen/Qwen3.5-2B
base_model_relation: quantized
library_name: llama.cpp
pipeline_tag: text-generation
tags:
  - gguf
  - llama.cpp
  - quantization
  - mixed-precision
---

# Qwen3.5 2B — TheStageAI GGUF

Four text-only GGUF checkpoints from **0.738 GB** to **2.013 GB**, evaluated on IFEval and MMLU-Pro where a complete release score is available.

[Qwen3.5 0.8B](https://huggingface.co/TheStageAI/Qwen3.5-0.8B-GGUF) · **[Qwen3.5 2B](https://huggingface.co/TheStageAI/Qwen3.5-2B-GGUF)** · [Qwen3.5 4B](https://huggingface.co/TheStageAI/Qwen3.5-4B-GGUF) · [Qwen3.5 9B](https://huggingface.co/TheStageAI/Qwen3.5-9B-GGUF) · [Gemma 4 E2B IT](https://huggingface.co/TheStageAI/gemma-4-E2B-it-GGUF) · [Gemma 4 E4B IT](https://huggingface.co/TheStageAI/gemma-4-E4B-it-GGUF) · [Gemma 4 12B IT](https://huggingface.co/TheStageAI/gemma-4-12B-it-GGUF)

## Choose a file

| Variant | Hub class | File | Size | Whole-file BPW | IFEval P / I (%) | MMLU-Pro (%) | Use |
| --- | --- | --- | ---: | ---: | ---: | ---: | --- |
| XS | `Q3_K_S` | `Qwen3.5-2B-XS-TS-Q3_K_S.gguf` | 0.738 GB | 3.140 | 52.68 / 64.15 | — | Minimum size |
| S | `Q4_K_S` | `Qwen3.5-2B-S-TS-Q4_K_S.gguf` | 0.967 GB | 4.109 | 63.22 / 73.38 | — | Compact |
| **M** | `Q4_K_M` | `Qwen3.5-2B-M-TS-Q4_K_M.gguf` | 1.073 GB | 4.562 | 66.54 / 75.54 | — | **Recommended · Uniform Q4** |
| L | `Q8_0` | `Qwen3.5-2B-L-TS-Q8_0.gguf` | 2.013 GB | 8.558 | 65.80 / 74.94 | — | Uniform Q8 |

> **Hub class versus file type:** XS and S are TheStage mixed-precision schedules. Their Hub labels describe the whole-file size class for discoverability; they are not stock `Q3_K_S`, `Q3_K_L`, or `Q4_K_S` conversions. M and L use the actual `MOSTLY_Q4_K_M` and `MOSTLY_Q8_0` GGUF file types. Exact tensor-type counts and SHA-256 digests are in [`release-manifest.json`](./release-manifest.json).

IFEval is shown as prompt-strict / instruction-strict. The BF16 reference scored **65.43 / 74.70** on IFEval. MMLU-Pro is not reported for the BF16 reference in this release.

## Quickstart

Use a current [llama.cpp](https://github.com/ggml-org/llama.cpp) build. The command below selects the recommended **M** file explicitly through its Hub quant label:

```bash
llama-cli -hf TheStageAI/Qwen3.5-2B-GGUF:Q4_K_M
```

## Evaluation

- **IFEval:** 541 prompts, native chat template, `enable_thinking=false`, temperature 0.
- **MMLU-Pro:** 12,032 questions for complete rows, native chat template, `enable_thinking=true`, temperature 1, top-p 0.95.
- A dash means **not reported**, not zero and not a score reconstructed from partial subjects.

These protocols intentionally exercise different operating modes. IFEval measures non-thinking instruction and format adherence; MMLU-Pro includes long sampled reasoning.

## What this release is

This release starts from [`Qwen/Qwen3.5-2B`](https://huggingface.co/Qwen/Qwen3.5-2B) at revision [`15852e8c16360a2fea060d615a32b45270f8a8fc`](https://huggingface.co/Qwen/Qwen3.5-2B/tree/15852e8c16360a2fea060d615a32b45270f8a8fc) and applies TheStageAI's scale-tuned deployment compression pipeline.

- **XS:** minimum-size adaptive mixed precision.
- **S:** compact adaptive mixed precision.
- **M:** uniform Q4 operating point.
- **L:** uniform Q8 operating point.

The four products are operating points, not a promise that benchmark scores increase monotonically with file size. The recommended row is selected separately for each base model from the release evaluations.

## Limitations

- The files in this repository contain the language-model GGUF. Multimodal projector files are not included.
- Small score reversals between BF16 and quantized rows should be read as evaluation variation, not as a claim that quantization improves the base model.
- For XS, use `enable_thinking=false`. A headline thinking-mode MMLU-Pro score is not reported when long generations do not produce a stable product metric.
- MMLU-Pro is not reported for this model in this release; diagnostic partial subjects are deliberately not aggregated.

## Provenance

The exact base revision, file sizes, GGUF file types, whole-file BPW, SHA-256 digests, held-out KL values, and evaluation IDs are recorded in [`release-manifest.json`](./release-manifest.json). Release filenames differ from the internal artifact paths; the payload bytes must match those digests exactly.

## License

The model weights are released under the upstream model's **apache-2.0** license. llama.cpp and any surrounding runtime code retain their own licenses.