Jais-3-32B-GGUF

This repository contains GGUF versions of inception42/Jais-3-32B for use with llama.cpp.

Multiple precision and quantization variants are available: BF16, Q8_0, Q6_K, Q5_K_M, Q5_0, Q4_K_M.

Deployment notes: K2 Horizon architecture support is merged into llama.cpp (ggml-org/llama.cpp#29535). Use a recent llama.cpp build, build it yourself from master, or wait for the next official llama.cpp release. Start llama-server with --jinja so that reasoning and tool calls are parsed.

Jais-3-32B is the large dense member of the Jais-3 family: a 32B-class decoder-only model with a 512K context window.

Other formats are available in the IFM/Jais-3 collection:

Jais-3-32B benchmark results

Benchmark Results

Reference models
Jais-3-32BFalcon-H1-34B-InstructGemma-4-31B-itMuse-Glimmer-30BQwen3.8-27BFanar-2-27B
# Params32B34B31B30B27B27B
# Activated params32B34B31B30B27B27B
ArchitectureDenseDenseDenseDenseDenseDense
Open-ended Generation
AraGen
AraGen v3 (3C3H)
09-2025
71.045.073.475.666.755.9
Correctness
81.053.285.487.777.165.7
Completeness
80.351.084.987.076.263.5
Conciseness
28.119.721.123.023.921.6
Helpfulness
76.445.479.781.271.559.0
Honesty
79.949.084.486.774.761.3
Harmlessness
80.351.484.987.776.564.4
Knowledge & Reasoning
General Knowledge & Comprehension
Arabic MMLU*
Accuracy
88.673.789.587.085.866.3
Belebele
Accuracy
95.086.695.994.394.461.8
STEM & Math
AraSTEM*
Accuracy
95.285.596.394.795.342.4
Arabic GSM8K*
Accuracy
94.390.293.795.293.287.7
GATmath*
Accuracy
98.563.498.698.598.487.8
Professional Domains
MedAraBench*
Accuracy
87.060.788.886.187.350.9
Arabic Legal (HELM)*
Accuracy
73.463.171.173.170.545.7
Arabic Language & Culture
Islamic Knowledge & Arab Culture
IslamicQA
Accuracy (mean of the 4 below)
88.771.189.783.777.152.0
PaLM-X 2025-2*
Accuracy
94.684.794.592.690.869.3
QIAS-1*
Accuracy
95.956.293.287.481.631.4
QIAS-2*
Accuracy
85.573.486.983.081.961.8
IslamicEval2025-1B*
Accuracy
78.870.184.071.954.145.5
Arab Culture
Accuracy (mean of the 3 below)
88.076.590.284.681.041.4
ArabCulture*
Accuracy
88.279.791.486.085.034.2
AraDiCE-Culture*
Accuracy
88.375.489.483.277.155.9
Arabic Cuisine^
Accuracy
87.474.489.884.781.034.0
Dialects
Dialectal Arabic MMLU*
Accuracy
88.571.589.487.087.028.2
Dialect Identification (QADI)
Accuracy
70.650.370.460.651.529.6
Alyah*
Accuracy
83.071.888.480.377.736.3
Linguistics & Literature
Arabic MMLU with diacritics
Accuracy
87.969.188.986.685.264.2
FannOrFlop-MCQ
Accuracy
71.862.472.668.362.546.4
Arabic Linguistics^
Accuracy (mean of the 2 below)
94.357.496.791.889.739.1
Diacritization
Accuracy
91.965.295.289.589.647.0
Parsing
Accuracy
96.749.798.394.189.831.3
Translation
FLORES200+
LLM-as-a-judge
91.570.692.884.881.871.2
Assistant Capabilities
Instruction Following
Arabic IFBench^
Accuracy (prompt-level strict)
88.732.280.884.281.340.0
Arabic IFEval
Accuracy
80.066.878.779.577.049.8
Long Context
AA-LCR (Ar)
Accuracy
75.014.066.0-79.0-
Factuality & Hallucination
HalluScore
Accuracy
67.342.469.769.154.641.7
SimpleQA Verified (Ar)
Aggregate (2/3 F-score + 1/3 non-hallucination rate)
28.17.510.613.38.011.0
F-score
7.16.610.718.29.210.4
Non-hallucination rate
70.19.410.53.75.412.2
Agentic
Terminal Bench 2.1
Accuracy (Avg@3)
63.71.143.551.779.81.1
GDPVal
Accuracy (@ 75% threshold)
75.50.030.548.278.20.0
MCPMark
Accuracy (Pass@1)
66.10.848.844.949.612.1
WildClawBench (English text-only subset)
Mean Score@3
42.010.744.143.252.110.3
Apex-Agents (420 text-only subset, loop agent)
Pass@1 (Accuracy)
25.20.712.422.128.60.2
GAIA v2 lit (Ar)
Accuracy (Pass@1)
51.59.742.436.455.812.1
Arabic File Processing
Accuracy (@ 75% threshold)
71.80.04.520.978.60.0
SWE-Bench Verified (Offline)
Accuracy (Pass@1)
72.40.459.071.473.50.4
BrowseComp-Plus
Accuracy
71.91.842.541.153.03.9

Starred (*) benchmarks have been audited for sample quality. Items with ambiguous questions, invalid answer options, or indefensible gold labels were excluded, since they can penalize a correct answer or reward a lucky guess. Scores on these benchmarks are reported on the cleaned sets. The audit methodology, the excluded samples and the verified benchmarks are available in the IFM/Arabic-Benchmark-Auditing collection.

Benchmarks marked with ^ will be released by IFM soon.

Scores above are for the BF16 source checkpoint. Quantized variants have not been separately evaluated yet.

Files

File Type Size
Jais-3-32B-BF16.gguf BF16 69.57 GB
Jais-3-32B-Q8_0.gguf Q8_0 36.97 GB
Jais-3-32B-Q6_K.gguf Q6_K 28.54 GB
Jais-3-32B-Q5_K_M.gguf Q5_K_M 24.70 GB
Jais-3-32B-Q5_0.gguf Q5_0 24.09 GB
Jais-3-32B-Q4_K_M.gguf Q4_K_M 21.08 GB

Quickstart

llama.cpp

cmake -B build -DGGML_CUDA=ON
cmake --build build -j --target llama-server
./build/bin/llama-server -hf IFM/Jais-3-32B-GGUF:Q8_0 --jinja -ngl 99 -c 32768 -np 1

Pick a variant with :BF16, :Q8_0, :Q6_K, :Q5_K_M, :Q5_0 or :Q4_K_M, or pass a local file with -m. --jinja enables the embedded chat template, which handles reasoning_effort and K2 Horizon tool calls; reasoning_content is returned separately from content.

API Usage

Recommended settings: reasoning_effort="high", the sampling hyperparameters in the table below, and at least 32,768 output tokens. Reasoning depth is selected per request through chat_template_kwargs. Thinking is returned in reasoning_content and the answer in content.

Parameter Default Thinking mode, English Thinking mode, Arabic
temperature 0.2 1.0 0.7
top_p 0.95 0.95 0.95
top_k 20 20 20
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8080/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="Jais-3-32B",
    messages=[{"role": "user", "content": "لدينا عددان مجموعهما 50 وحاصل ضربهما 600. ما هما العددان؟ اشرح خطوات الحل."}],
    temperature=0.7,
    top_p=0.95,
    max_tokens=32768,
    extra_body={"top_k": 20, "chat_template_kwargs": {"reasoning_effort": "high"}},
)
message = response.choices[0].message
print("Reasoning:", getattr(message, "reasoning_content", None))
print("Answer:", message.content)
import json
from datetime import datetime
from zoneinfo import ZoneInfo
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8080/v1", api_key="EMPTY")
TIMEZONES = {"أبوظبي": "Asia/Dubai", "الرياض": "Asia/Riyadh", "القاهرة": "Africa/Cairo", "لندن": "Europe/London"}

def get_current_time(city: str) -> dict:
    if city not in TIMEZONES:
        return {"error": f"مدينة غير معروفة: {city}"}
    return {"city": city, "time": datetime.now(ZoneInfo(TIMEZONES[city])).strftime("%H:%M")}

tools = [{
    "type": "function",
    "function": {
        "name": "get_current_time",
        "description": "يعيد الوقت الحالي في مدينة معينة",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string", "description": "اسم المدينة بالعربية"}},
            "required": ["city"],
        },
    },
}]
available_tools = {"get_current_time": get_current_time}
messages = [{"role": "user", "content": "كم الساعة الآن في القاهرة؟"}]
for _ in range(5):  # call the model, run any requested tools, repeat until it answers
    message = client.chat.completions.create(
        model="Jais-3-32B",
        messages=messages,
        tools=tools,
        temperature=0.7,
        top_p=0.95,
        max_tokens=32768,
        extra_body={
            "top_k": 20,
            "chat_template_kwargs": {"reasoning_effort": "high", "tool_call_format": "xml"},
        },
    ).choices[0].message
    messages.append(message)
    if not message.tool_calls:
        break
    for call in message.tool_calls:
        tool = available_tools.get(call.function.name)
        arguments = json.loads(call.function.arguments)
        result = tool(**arguments)
        messages.append({"role": "tool", "tool_call_id": call.id,
                         "content": json.dumps(result, ensure_ascii=False)})

print(message.content)

Our model supports multiple tool-call formats, which can be changed with chat_template_kwargs. The supported values are json, xml, and xml_typed. The default is xml. Keep --tool-call-parser k2_horizon enabled to parse the selected format.

Best Practices

  1. Reasoning effort: always high. All reported results use high reasoning effort. Pass {"chat_template_kwargs": {"reasoning_effort": "high"}} on every request. No reasoning ({"reasoning_effort": "low"}) is not recommended except for research on reasoning efforts.
  2. Sampling parameters. Use the values in the table above: temperature=0.2, top_p=0.95, top_k=20 by default. In thinking mode, set temperature=1.0 for English and temperature=0.7 for Arabic. Keep min_p=0.0, presence_penalty=0.0 and repetition_penalty=1.0.
  3. Parsers. With --jinja, llama.cpp detects the K2 Horizon template and parses reasoning and tool calls itself; no parser flags are needed.

Intended Use and Limitations

Intended use: The model is designed for Arabic and English assistant and agentic applications, including reasoning and tool-use tasks, as well as for research and experimentation.

Out of scope: The model should not be used in ways that violate applicable laws or regulations. The following uses are also considered out of scope:

  • Harmful or malicious use: generating hateful, extremist, discriminatory, or deliberately misleading content, or using the model to support illegal or harmful activities.
  • Sensitive information: using the model to generate, expose, or infer personal, confidential, or otherwise sensitive information about individuals or organizations.
  • Unsupported languages: applications that depend on reliable performance in languages other than Arabic or English.
  • High-stakes decisions: relying on the model to make medical, legal, financial, or safety-critical decisions without appropriate human oversight.

Limitations: The model can produce inaccurate or made-up content and may reflect biases present in its training data. Evaluate it for your use case and add safeguards before deploying it.

Citation

@misc{ifm2026jais3introducingarabiccentric,
      title={Jais 3: Introducing Arabic-centric Agentic AI},
      author={Team IFM and Haonan Li and Hadi Abdine and Nikhil Ranjan and Mohamed Anwar and Ahmad Chamma and Amr Mohamed and Ahmed Frikha and Dani Bouch and Maaz Maqsood and George Ibrahim and Mostafa Awad and Nurkhan Laiyk and Hasan Iqbal and Abhishek Maiti and Bilal Bouardi and Shivam Chauhan and Ahmer Tabassum and Owais Aijaz and Omar Arbi and Neha Sengupta and Aaryamonvikram Singh and Abdelhakim Freihat and Sarath Chandran and Rahul Pal and Mohammed Bellam and Sajid Siddiki and Sarah AlBarri and Debopriyo Banerjee and Ali El Filali and Samta Kamboj and Rifo Genadi and Ahmed Attia and Go Inoue and Shuonan Dong and Amine Khelif Khelif and Evan Dufraisse and Florian Valade and Tapas Sen and Guowei He and Sunil Sahu and Onkar Pandit and Kareem Magdy and Sheikh Khalid bin Khalifa Al Nahyan and Samujjwal Ghosh and Zainul Abedien and Ahmed Quraishi and Larry Murray and Avraham Sheinin and Natalia Vassilieva and Preslav Nakov and Guokan Shang and Xuguang Hector Ren and Zhengzhong Liu and Eric Xing},
      year={2026},
      eprint={2611.10224},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2611.10224},
}
Downloads last month
4,174
GGUF
Model size
35B params
Architecture
k2-horizon
Hardware compatibility
Log In to add your hardware

4-bit

5-bit

6-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for IFM/Jais-3-32B-GGUF

Quantized
(8)
this model

Collection including IFM/Jais-3-32B-GGUF

Paper for IFM/Jais-3-32B-GGUF