README.md

---
license: apache-2.0
base_model:
  - Qwen/Qwen3.5-4B
base_model_relation: quantized
library_name: llama.cpp
pipeline_tag: text-generation
thumbnail: https://huggingface.co/TheStageAI/Qwen3.5-4B-GGUF/resolve/main/assets/thestage-edge-models-header.png
tags:
  - gguf
  - llama.cpp
  - quantized
  - mixed-precision
  - local-inference
  - qwen3.5
---

<p align="center">
  <img src="./assets/thestage-edge-models-header.png" width="100%" alt="TheStageAI Edge Models: the right model at every memory budget">
</p>

<h1 align="center">Qwen3.5 4B</h1>

<p align="center">
  Four GGUF checkpoints for llama.cpp, from 1.52 GB to 4.49 GB.
  <br>
  <strong>Start with M: 2.39 GB and 98.4% of the BF16 instruction-strict IFEval score.</strong>
</p>

<p align="center">
  <a class="inline-block" href="https://github.com/TheStageAI/edge-lm"><img class="dark:hidden" src="./assets/cta-edge-lm-light.svg" width="146" height="42" alt="Explore edge-lm on GitHub"><img class="hidden dark:block" src="./assets/cta-edge-lm-dark.svg" width="146" height="42" alt="Explore edge-lm on GitHub"></a>&nbsp;
  <a class="inline-block" href="https://docs.thestage.ai/"><img class="dark:hidden" src="./assets/cta-docs-light.svg" width="120" height="42" alt="Read TheStageAI documentation"><img class="hidden dark:block" src="./assets/cta-docs-dark.svg" width="120" height="42" alt="Read TheStageAI documentation"></a>&nbsp;
  <a class="inline-block" href="https://app.thestage.ai/"><img class="dark:hidden" src="./assets/cta-platform-light.svg" width="146" height="42" alt="Open TheStageAI Platform"><img class="hidden dark:block" src="./assets/cta-platform-dark.svg" width="146" height="42" alt="Open TheStageAI Platform"></a>
</p>

## Choose a checkpoint

| Tier | Size | Best for | File |
| --- | ---: | --- | --- |
| XS | 1.52 GB | Minimum footprint | [Download](./Qwen3.5-4B-XS-TS-Q3_K_S.gguf) |
| S | 1.90 GB | Compact | [Download](./Qwen3.5-4B-S-TS-Q4_K_S.gguf) |
| **M** | **2.39 GB** | **Recommended** | [Download](./Qwen3.5-4B-M-TS-Q4_K_M.gguf) |
| L | 4.49 GB | High-precision Q8 | [Download](./Qwen3.5-4B-L-TS-Q8_0.gguf) |

**Other Qwen 3.5 sizes:** <a href="https://huggingface.co/TheStageAI/Qwen3.5-0.8B-GGUF">0.8B</a> &nbsp;·&nbsp; <a href="https://huggingface.co/TheStageAI/Qwen3.5-2B-GGUF">2B</a> &nbsp;·&nbsp; <a href="https://huggingface.co/TheStageAI/Qwen3.5-9B-GGUF">9B</a>

## Quickstart

```bash
llama-cli \
  --hf-repo TheStageAI/Qwen3.5-4B-GGUF \
  --hf-file Qwen3.5-4B-M-TS-Q4_K_M.gguf
```

## Why we recommend M

At 2.39 GB, M retains 98.4% of the BF16 instruction-strict IFEval score. On MMLU-Pro, it retains 99.1% of the BF16 score. It uses 47% less disk than L.

| Tier | IFEval P / I (%) | MMLU-Pro (%) |
| --- | ---: | ---: |
| BF16 reference | 82.44 / 87.53 | 79.55 |
| L | 81.70 / 87.05 | 79.59 |
| **M** | 80.22 / 86.09 | 78.86 |
| S | 77.82 / 83.93 | 74.39 |
| XS | 70.43 / 78.30 | — |

P / I means prompt-strict / instruction-strict. IFEval uses deterministic non-thinking decoding; MMLU-Pro uses sampled long-form reasoning. Only complete model-level scores are reported. A dash means not reported.

> **Reasoning:** XS is intended for non-thinking use. Run it with `--reasoning off`. Choose S, M, or L for long-form reasoning.

<details>
<summary><b>Evaluation protocol</b></summary>

- **IFEval:** 541 prompts, native chat template, `enable_thinking=false`, temperature 0.
- **MMLU-Pro:** 12,032 questions, vLLM, 0-shot, native chat template `qwen_mc_json_v1`, `enable_thinking=true`; `temperature=1`, `top_p=0.95`, `top_k=20`, `min_p=0`, `presence_penalty=1.5`, `frequency_penalty=0`, `repetition_penalty=1`, `seed=42`; `max_model_len=40960`, `max_new_tokens=32768`; dataset revision `b189ec765aa7ed75c8acfea42df31fdae71f97be`.
- The headline BF16 comparison uses instruction-strict IFEval; the raw scores are shown in the table.

In matched long-form reasoning diagnostics, XS produced longer trajectories and reached the 32,768-token output limit more often than S. The XS MMLU-Pro cell is marked `—` for that reason.

</details>

## How the checkpoints are built

All four checkpoints use the same production PTQ pipeline. The precision map is the only tier-specific part.

### 1. Fit native GGUF codes

Calibration activations produce a curvature objective weighted by true Fisher information for each quantized projection. We adapt the scale and minimum initialization from [NeUQI](https://arxiv.org/abs/2505.17595) to that objective, then solve integer codes on the target GGUF grid with a guarded cyclic coordinate-descent solver inspired by [QuantEase](https://arxiv.org/abs/2309.01885). A final K-quant pass tunes stored scales and minima while keeping packed codes fixed.

### 2. Reconstruct the deployed trajectory

Layers are calibrated in execution order against activations from the already-quantized prefix. A dense reference path measures accumulated drift. [Quantization Error Propagation (QEP)](https://arxiv.org/abs/2504.09629) adds that drift to the next reconstruction target, so later layers optimize for the inputs they receive at inference time.

### 3. Allocate the byte budget

XS and S can choose Q2_K through Q8_0 for each quantizable group. The optimizer trades changes in the teacher distribution against the actual encoded byte cost, including scale and minimum metadata. [ANNA](https://docs.thestage.ai/qlip/docs/source/anna_api.html) provides constrained configuration search. [RCO](https://arxiv.org/abs/2605.00649) provides an exact-budget route ([code](https://github.com/IST-DASLab/RCO)). For this model, RCO selected both the XS and S precision maps. M and L keep the decoder qtypes at Q4_K_M and Q8_0, respectively, and use the same reconstruction and scale-tuning stages.

After schedule selection, PTQ runs again from the source weights. Each layer is then calibrated with the final upstream precision choices.

### 4. Align the full model

A short affine distillation pass tunes native FP16 scales and minima while qtypes, packed codes, dense weights, and tensor layouts stay fixed. The loss matches the teacher's next-token distribution without changing file size or runtime layout.

We load-test the shipping GGUF and evaluate it on a held-out set of 3,072 sequences with next-token KL. [`release-manifest.json`](./release-manifest.json) records its SHA-256, downstream evaluation IDs, and tensor metadata. Final recommendations use complete-model benchmarks.

<details>
<summary><b>File details</b></summary>

| Tier | Hub selector | GGUF file type | Whole-file BPW |
| --- | --- | --- | ---: |
| XS | `Q3_K_S` | `MOSTLY_Q2_K` | 2.889 |
| S | `Q4_K_S` | `MOSTLY_Q2_K` | 3.622 |
| M | `Q4_K_M` | `MOSTLY_Q4_K_M` | 4.538 |
| L | `Q8_0` | `MOSTLY_Q8_0` | 8.533 |

The Hub selector controls sidebar grouping and download discovery. For XS and S it approximates the whole-file size class; [`release-manifest.json`](./release-manifest.json) contains the exact tensor mix. M and L keep their decoder qtypes at Q4_K_M and Q8_0 within the same production PTQ pipeline.

Runtime memory also includes KV cache and buffers, which grow with context length.

</details>

## TheStageAI deployment stack

These GGUF files target llama.cpp-compatible runtimes. [edge-lm](https://github.com/TheStageAI/edge-lm) runs compressed MLX models on Macs and iPhones. [ANNA](https://docs.thestage.ai/qlip/docs/source/anna_api.html) searches compression configurations under size or compute constraints. The [TheStageAI Platform](https://app.thestage.ai/) and [documentation](https://docs.thestage.ai/) cover compression, compilation, and serving workflows.

For a specific device, latency target, or memory budget, [talk to our team](https://app.thestage.ai/contact).

## Reproducibility

- **Release:** July 21, 2026.
- **Base model:** [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B) at revision [`851bf6e8`](https://huggingface.co/Qwen/Qwen3.5-4B/tree/851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a).
- **Manifest:** [`release-manifest.json`](./release-manifest.json) records the exact base revision, byte sizes, GGUF file types, whole-file BPW, tensor inventories, SHA-256 digests, held-out KL values, and evaluation IDs.
- **Runtime gate:** export and load checks used [llama.cpp revision `bec4772f`](https://github.com/ggml-org/llama.cpp/commit/bec4772f6a2527d371557b5d2032641e5ff7619c).

## Citation

If you use this checkpoint, cite this release and follow the upstream model's citation guidance:

```bibtex
@misc{thestageai2026qwen3p54bgguf,
  author       = {{TheStageAI}},
  title        = {Qwen3.5 4B: TheStageAI GGUF Release},
  year         = {2026},
  month        = {jul},
  howpublished = {Hugging Face model release},
  url          = {https://huggingface.co/TheStageAI/Qwen3.5-4B-GGUF},
  note         = {XS, S, M, and L deployment tiers},
}
```

<details>
<summary><b>References</b></summary>

- [ANNA](https://docs.thestage.ai/qlip/docs/source/anna_api.html), TheStageAI's constrained compression configuration search.
- [RCO: Model Compression with Exact Budget Constraints via Riemannian Manifolds](https://arxiv.org/abs/2605.00649) ([code](https://github.com/IST-DASLab/RCO)).
- [NeUQI: Near-Optimal Uniform Quantization Parameter Initialization for Low-Bit LLMs](https://arxiv.org/abs/2505.17595).
- [QuantEase: Optimization-based Quantization for Language Models](https://arxiv.org/abs/2309.01885).
- [Quantization Error Propagation: Revisiting Layer-Wise Post-Training Quantization](https://arxiv.org/abs/2504.09629).

</details>

## License

The checkpoint weights use the upstream model's **Apache-2.0** license. llama.cpp and other runtime software keep their own licenses.