license: apache-2.0 base_model: inception42/Jais-3-7B base_model_relation: quantized library_name: gguf pipeline_tag: text-generation language:

• ar
• en tags:
• sakura
• gguf
• imatrix
• jais-3
• k2-horizon
• arabic
• uncensored
• abliterated
• not-for-all-audiences

For adults only; you are responsible for what you generate and for the laws that apply to you. The language weights of this model were edited to refuse less; it still refuses some of the harmful prompts we tested (4 to 6 of 104), and we measured only English prompts in a single setting (see below). Do not use it to harm others.

Sakura — Jais-3-7B, uncensored (GGUF)

Jais-3-7B (Arabic and English, dense, K2 Horizon architecture) in a version whose weights were edited by us with an abliteration run so that it refuses less, as GGUF with importance-matrix quantizations. The unmodified model is in Sakura-Jais-3-7B-GGUF. This is version 2; the first version of these files refused much more and was replaced.

What was changed

• The language weights were edited with an abliteration method (a refusal direction is removed from part of the weight matrices) on the original checkpoint inception42/Jais-3-7B (revision 94d69f49), with settings chosen and tuned by us; all other weights are unchanged.
• The files were made from the original BF16 weights with llama.cpp and quantized with an importance matrix computed by us on the edited model (English and German calibration text, no Arabic).

Files

Run

llama-server -m Sakura-Jais-3-7B-Uncensored-Q5_K_M-*.gguf --jinja -ngl 99 -c 8192

Jais-3 is a model that always thinks: its chat template opens <ifm|think>, <ifm|think_fast> or <ifm|think_faster> for reasoning_effort high (default), medium or low. Pass the effort per request as chat_template_kwargs: {"reasoning_effort": "low"}. The architecture (K2 Horizon) needs a llama.cpp build with the support merged on 2026-10-06 (PR 29535); we used b11594 and b11681, older builds do not know it.

What we measured

Refusals. The 104 harmful prompts of the test split of mlabonne/harmful_behaviors, plain chat without a system prompt, reasoning_effort low, temperature 0, up to 1500 tokens, llama-server b11594; the answer after the thinking block is counted; a refusal is an English refusal phrase or one of 17 Arabic refusal phrases (the models often answer English prompts in Arabic). These prompts were not used to compute the edit, but they are not held out from our choice of settings (we looked at these 104 prompts when choosing).

Language: with these English prompts, 15 of 104 answers of this version's Q8_0 and 52 of 104 of its Q4_K_M are in Arabic (unmodified Q4_K_M: 30); we do not know why the Q4_K_M files switch language so often. An answer that is not counted as a refusal has not necessarily done what was asked, and we did not check the quality of the answers beyond reading some. We did not repeat the measurement at reasoning_effort high for this version (the first version had 13 of 104 at high with 21 answers cut off).

Distance to the unmodified model (Q8_0 of each version against the unmodified Q8_0; KL divergence, llama-perplexity b11594, 20 chunks of 2048 tokens of an English and a web text):

Quantization loss (KL divergence of each file against the Q8_0 of this repository, same protocol; the reference is Q8_0, not BF16):

Not measured: anything in Arabic (quality, refusals, perplexity), any capability benchmark, refusals with a system prompt, other prompt sets, multi-turn chats, speed, the high reasoning effort for this version.

License

Apache-2.0 as the original model (LICENSE). The files are modified (edited and quantized) versions of the original weights. See NOTICE.

Credits

• Model: inception42/Jais-3-7B.
• Conversion, quantization and measurement: ggml-org/llama.cpp (MIT).

中文版见 README_zh.md。