---

license: mit

license_link: LICENSE

base_model: microsoft/Fara1.5-4B

base_model_relation: quantized

library_name: gguf

pipeline_tag: image-text-to-text

tags:

- sakura

- sakura-mini

- gguf

- llama.cpp

- quantization

- imatrix

- mixed-quant

- qwen3_5

---

Sakura — Fara1.5-4B (GGUF, three sizes)

<img src="assets/sakura-logo.svg" alt="Sakura logo" width="560" />

Fara1.5-4B is the 4B member of Microsoft's computer-use agent family for web browsing (Qwen3.5 base, multimodal: it reads browser screenshots and emits structured tool calls). This repository holds three GGUF files of different size made from the BF16 weights with our measured mixed-codec method. Pick one on the Files page; the table below says what each file is.

Independent community quantization, not an official Microsoft release. Part of the Sakura Mini line.

Quality at a glance

Same method, same texts, same machine for every file; lower KL divergence means closer to the BF16 original:

• Sakura-Fara1.5-4B-1.89GiB.gguf vs bartowski IQ3_XXS (1.97 GiB): KL divergence to the BF16 original 40 % lower on average over all 3 texts (0.08 GiB smaller).
• Sakura-Fara1.5-4B-2.29GiB.gguf vs bartowski IQ4_XS (2.37 GiB): KL divergence to the BF16 original 14 % lower on average over all 3 texts (0.08 GiB smaller).
• Sakura-Fara1.5-4B-2.66GiB.gguf vs bartowski Q4_K_M (2.69 GiB): KL divergence to the BF16 original 42 % lower on average over all 3 texts (0.03 GiB smaller).

Why: the bit budget per tensor is allocated from measured sensitivity (see How it was made), not from fixed rules. Full table below.

Which file should I take?

| File | For |

|---|---|

| Sakura-Fara1.5-4B-1.89GiB.gguf (1.88 GiB, 3.83 bpw) | smallest: for tight memory budgets; expect visible quality loss |

| Sakura-Fara1.5-4B-2.29GiB.gguf (2.28 GiB, 4.65 bpw) | middle: best balance of size and quality for most machines |

| Sakura-Fara1.5-4B-2.66GiB.gguf (2.65 GiB, 5.42 bpw) | largest: closest to the BF16 original of the three |

The files

| File | Size | bits/weight | KLD en | KLD dev | KLD wiki | same top token (avg) | Tensor types by size |

|---|---:|---:|---:|---:|---:|---:|---|

| Sakura-Fara1.5-4B-1.89GiB.gguf | 1.88 GiB | 3.83 | 0.0968 | 0.0898 | 0.1496 | 86.9 % | IQ4_XS 35%, IQ3_S 33%, Q3_K 14%, Q4_K 12%, IQ3_XXS 3%, Q5_K 2% |

| Sakura-Fara1.5-4B-2.29GiB.gguf | 2.28 GiB | 4.65 | 0.0332 | 0.0303 | 0.0484 | 92.6 % | IQ4_XS 63%, Q5_K 36% |

| Sakura-Fara1.5-4B-2.66GiB.gguf | 2.65 GiB | 5.42 | 0.0136 | 0.0160 | 0.0216 | 95.2 % | Q5_K 92%, IQ4_XS 7% |

"Tensor types by size" lists the share of the file's bytes per ggml type (all are standard llama.cpp types, so the files run in stock llama.cpp). Quality columns: mean KL divergence of the quantized model's next-token distribution against the BF16 source on held-out text (lower is better) and how often the most likely token is unchanged (higher is better).

• mmproj-Fara1.5-4B-f16.gguf: vision projector (unchanged from bartowski's F16 conversion).

Comparison at similar size

Same texts, same context, same chunks, measured by us with llama-perplexity (BF16 source = reference; PPL of the BF16 model: en: 2.169, dev: 2.242, wiki: 10.998). The reference quants are: bartowski (imatrix), measured the same way. Sorted by size.

| Quant | Source | Size | KLD en | KLD dev | KLD wiki | PPL en | same top token (avg) |

|---|---|---:|---:|---:|---:|---:|---:|

| Sakura-Fara1.5-4B-1.89GiB.gguf | this repository | 1.89 GiB | 0.0968 | 0.0898 | 0.1496 | 2.262 | 86.9 % |

| IQ3_XXS | bartowski (imatrix) | 1.97 GiB | 0.1590 | 0.1526 | 0.2436 | 2.368 | 84.7 % |

| Sakura-Fara1.5-4B-2.29GiB.gguf | this repository | 2.29 GiB | 0.0332 | 0.0303 | 0.0484 | 2.201 | 92.6 % |

| IQ4_XS | bartowski (imatrix) | 2.37 GiB | 0.0425 | 0.0364 | 0.0505 | 2.211 | 92.2 % |

| Sakura-Fara1.5-4B-2.66GiB.gguf | this repository | 2.66 GiB | 0.0136 | 0.0160 | 0.0216 | 2.177 | 95.2 % |

| Q4_K_M | bartowski (imatrix) | 2.69 GiB | 0.0257 | 0.0266 | 0.0351 | 2.186 | 93.9 % |

| Q5_K_M | bartowski (imatrix) | 3.09 GiB | 0.0086 | 0.0077 | 0.0103 | 2.173 | 96.8 % |

Method: 12 chunks of 512 tokens per text; texts: English held-out text, developer-text held-out set, English encyclopedia held-out text (not used in any calibration). Single measurement on one machine; small KLD differences at similar size are not a quality ranking.

How it was made

1. BF16 GGUF of the original model and the importance matrix come from bartowski/Fara1.5-4B-GGUF; we used both unchanged (credit and thanks to bartowski).
2. Every large weight matrix was quantized once per candidate type (Q2_K, IQ2_S, IQ3_XXS, IQ3_S, Q3_K, IQ4_XS, Q4_K, Q5_K, Q6_K, Q8_0) with llama-quantize and the importance matrix; the error of each choice was estimated per matrix (importance-weighted) and scaled by a sensitivity factor measured with real KL divergence: every group of tensors (for example the FFN down projections, the attention value projections, the first and last layers) was moved alone to a lower-bit type, and the KL divergence it caused was compared with its quantization error.
3. An exact budget allocation (multiple-choice knapsack over bytes) picked one type per matrix for each target size; the final model was assembled from the stored tensors without re-quantizing.

We do not publish per-tensor choices. Norms, small tensors and the multi-token-prediction layers stay at high precision.

Limits

• Measured only with KL divergence and perplexity on short held-out texts and not with downstream benchmarks; do not read it as a task-quality claim.
• Quantization always costs some quality; the smallest file costs the most.
• Fara is a computer-use agent. Microsoft recommends running it only inside a sandbox with monitoring (MagenticLite) and describes safety behaviour (stopping and asking before payments, personal data and irreversible actions). Quantization can change behaviour; we did not test agent safety or screenshot grounding, only text perplexity/KLD.
• The vision projector mmproj-Fara1.5-4B-f16.gguf is bartowski's conversion (unchanged, F16); it is needed for screenshots and is not quantized by us.

Credits

• Original model: Fara1.5-4B (Microsoft, MIT), MIT license (included as LICENSE).
• BF16 GGUF, importance matrix and the reference quants: bartowski.
• GGUF format and tools: ggml-org/llama.cpp (MIT).

---

<details><summary>中文说明 · 樱花 (Simplified Chinese)</summary>

> English above. 本节为上文的中文翻译(Sakura = 樱花 yīnghuā);完整的独立中文版见 README_zh.md。

Sakura — Fara1.5-4B(GGUF,三种大小)

<img src="assets/sakura-logo.svg" alt="Sakura logo" width="560" />

Fara1.5-4B 是 Microsoft 面向网页浏览的计算机使用(computer-use)智能体系列中的 4B 成员(基于 Qwen3.5,多模态:读取浏览器截图并输出结构化的工具调用)。本仓库包含 三个不同大小的 GGUF 文件,由 BF16 权重通过我们实测得出的混合编码方法制作。请在 Files 页面选择其中一个;下表说明了每个文件的用途。

这是独立的社区量化版本,不是 Microsoft 的官方发布。属于 Sakura Mini 系列。

质量概览

每个文件使用相同的方法、相同的文本、同一台机器;KL 散度越低,越接近 BF16 原始模型:

• Sakura-Fara1.5-4B-1.89GiB.gguf 对比 bartowski IQ3_XXS (1.97 GiB):相对于 BF16 原始模型,在全部 3 个文本上平均 KL 散度 低 40 %(小 0.08 GiB)。
• Sakura-Fara1.5-4B-2.29GiB.gguf 对比 bartowski IQ4_XS (2.37 GiB):相对于 BF16 原始模型,在全部 3 个文本上平均 KL 散度 低 14 %(小 0.08 GiB)。
• Sakura-Fara1.5-4B-2.66GiB.gguf 对比 bartowski Q4_K_M (2.69 GiB):相对于 BF16 原始模型,在全部 3 个文本上平均 KL 散度 低 42 %(小 0.03 GiB)。

原因:每个张量的比特预算是根据实测的敏感度分配的(见 制作方式),而不是按固定规则。完整表格见下文。

我该选哪个文件?

| 文件 | 适用场景 |

|---|---|

| Sakura-Fara1.5-4B-1.89GiB.gguf(1.88 GiB,3.83 bpw) | 最小:适合内存预算紧张的情况;预计质量会有明显损失 |

| Sakura-Fara1.5-4B-2.29GiB.gguf(2.28 GiB,4.65 bpw) | 居中:对大多数机器而言,大小与质量的最佳平衡 |

| Sakura-Fara1.5-4B-2.66GiB.gguf(2.65 GiB,5.42 bpw) | 最大:三者中最接近 BF16 原始模型 |

这些文件

| 文件 | 大小 | bits/weight | KLD en | KLD dev | KLD wiki | same top token (avg) | 各张量类型(按大小) |

|---|---:|---:|---:|---:|---:|---:|---|

| Sakura-Fara1.5-4B-1.89GiB.gguf | 1.88 GiB | 3.83 | 0.0968 | 0.0898 | 0.1496 | 86.9 % | IQ4_XS 35%, IQ3_S 33%, Q3_K 14%, Q4_K 12%, IQ3_XXS 3%, Q5_K 2% |

| Sakura-Fara1.5-4B-2.29GiB.gguf | 2.28 GiB | 4.65 | 0.0332 | 0.0303 | 0.0484 | 92.6 % | IQ4_XS 63%, Q5_K 36% |

| Sakura-Fara1.5-4B-2.66GiB.gguf | 2.65 GiB | 5.42 | 0.0136 | 0.0160 | 0.0216 | 95.2 % | Q5_K 92%, IQ4_XS 7% |

“各张量类型(按大小)”列出每个 ggml 类型占文件字节数的比例(全部是标准的 llama.cpp 类型,因此这些文件可在原版 llama.cpp 中运行)。质量列:量化模型的下一个 token 分布相对于 BF16 源文件在留出文本上的平均 KL 散度(越低越好),以及最可能的 token 保持不变的频率(越高越好)。

• mmproj-Fara1.5-4B-f16.gguf:视觉投影器(与 bartowski 的 F16 转换版本相同,未改动)。

相近大小的比较

相同的文本、相同的上下文、相同的分块,由我们使用 llama-perplexity 测量(BF16 源文件 = 参照;BF16 模型的 PPL:en: 2.169,dev: 2.242,wiki: 10.998)。参照量化为:bartowski(imatrix),以相同方式测量。按大小排序。

| 量化 | 来源 | 大小 | KLD en | KLD dev | KLD wiki | PPL en | same top token (avg) |

|---|---|---:|---:|---:|---:|---:|---:|

| Sakura-Fara1.5-4B-1.89GiB.gguf | 本仓库 | 1.89 GiB | 0.0968 | 0.0898 | 0.1496 | 2.262 | 86.9 % |

| IQ3_XXS | bartowski (imatrix) | 1.97 GiB | 0.1590 | 0.1526 | 0.2436 | 2.368 | 84.7 % |

| Sakura-Fara1.5-4B-2.29GiB.gguf | 本仓库 | 2.29 GiB | 0.0332 | 0.0303 | 0.0484 | 2.201 | 92.6 % |

| IQ4_XS | bartowski (imatrix) | 2.37 GiB | 0.0425 | 0.0364 | 0.0505 | 2.211 | 92.2 % |

| Sakura-Fara1.5-4B-2.66GiB.gguf | 本仓库 | 2.66 GiB | 0.0136 | 0.0160 | 0.0216 | 2.177 | 95.2 % |

| Q4_K_M | bartowski (imatrix) | 2.69 GiB | 0.0257 | 0.0266 | 0.0351 | 2.186 | 93.9 % |

| Q5_K_M | bartowski (imatrix) | 3.09 GiB | 0.0086 | 0.0077 | 0.0103 | 2.173 | 96.8 % |

方法:每个文本取 12 个块,每块 512 个 token;文本为:英文留出文本、开发者文本留出集、英文百科留出文本(未用于任何校准)。这是在一台机器上的单次测量;相近大小下 KLD 的细小差异并不构成质量排名。

制作方式

1. 原始模型的 BF16 GGUF 和重要性矩阵来自 bartowski/Fara1.5-4B-GGUF;我们原样使用了这两者(感谢并归功于 bartowski)。
2. 每个大型权重矩阵针对每种候选类型(Q2_K、IQ2_S、IQ3_XXS、IQ3_S、Q3_K、IQ4_XS、Q4_K、Q5_K、Q6_K、Q8_0)使用 llama-quantize 和重要性矩阵各量化一次;每种选择的误差按矩阵估计(按重要性加权),并乘以一个 用真实 KL 散度测得的敏感度系数:每组张量(例如 FFN down 投影、注意力 value 投影、首尾几层)被单独移到较低比特的类型,并将它造成的 KL 散度与其量化误差进行比较。
3. 精确的预算分配(以字节为单位的多选背包问题)为每个目标大小为每个矩阵挑选一种类型;最终模型由已存储的张量组装而成,无需重新量化。

我们不公布逐张量的选择。归一化层、小张量和多 token 预测层保持高精度。

局限

• 仅用短的留出文本上的 KL 散度和困惑度测量,没有使用下游基准测试;请勿将其理解为任务质量的声明。
• 量化总会损失一些质量;最小的文件损失最大。
• Fara 是一个 计算机使用智能体。Microsoft 建议仅在带监控的沙箱(MagenticLite)中运行它,并描述了其安全行为(在涉及付款、个人数据和不可逆操作之前停下并询问)。量化可能改变行为;我们 没有 测试智能体的安全性或截图定位能力,只测试了文本困惑度/KLD。
• 视觉投影器 mmproj-Fara1.5-4B-f16.gguf 是 bartowski 的转换版本(未改动,F16);截图需要它,我们没有对它进行量化。

致谢

• 原始模型:Fara1.5-4B (Microsoft, MIT),MIT 许可证(以 LICENSE 随附)。
• BF16 GGUF、重要性矩阵和参照量化:bartowski。
• GGUF 格式和工具:ggml-org/llama.cpp (MIT)。

</details>