Jais-3-375B-A23B

Jais-3-375B-A23B is the flagship of the Jais-3 family: a sparse Mixture-of-Experts model that stores 375B parameters and runs 23B per token, with a 512K context window.

An FP8-quantized version is available at IFM/Jais-3-375B-A23B-FP8.

Jais-3-375B-A23B benchmark results

Benchmark Results

Reference models
Jais-3-375B-A23BQwen3.5-397B-A17BMiniMax M3Qwen3.8-Flash-NextNVIDIA Nemotron 3 Ultra
# Params375B397B428B180B550B
# Activated params23B17B23B6B55B
ArchitectureMoEMoEMoEMoEMoE
Open-ended Generation
AraGen
AraGen v3 (3C3H)
09-2025
72.478.971.974.972.8
Correctness
82.190.786.686.286.1
Completeness
80.990.084.685.685.8
Conciseness
31.527.514.324.416.6
Helpfulness
78.285.775.081.577.4
Honesty
80.789.584.985.885.0
Harmlessness
81.190.385.985.985.8
Knowledge & Reasoning
General Knowledge & Comprehension
Arabic MMLU*
Accuracy
89.389.186.489.388.1
Belebele
Accuracy
95.694.892.395.593.1
STEM & Math
AraSTEM*
Accuracy
95.596.695.696.395.2
Arabic GSM8K*
Accuracy
86.893.693.294.493.7
GATmath*
Accuracy
98.498.697.898.598.4
Professional Domains
MedAraBench*
Accuracy
89.590.887.790.888.4
Arabic Legal (HELM)*
Accuracy
71.371.674.969.577.0
Arabic Language & Culture
Islamic Knowledge & Arab Culture
IslamicQA
Accuracy (mean of the 4 below)
90.288.083.784.082.1
PaLM-X 2025-2*
Accuracy
95.394.692.894.793.5
QIAS-1*
Accuracy
95.486.180.491.483.6
QIAS-2*
Accuracy
87.786.885.187.884.8
IslamicEval2025-1B*
Accuracy
82.384.476.661.966.7
Arab Culture
Accuracy (mean of the 3 below)
89.187.284.786.687.9
ArabCulture*
Accuracy
89.889.186.689.889.1
AraDiCE-Culture*
Accuracy
87.783.882.282.788.8
Arabic Cuisine^
Accuracy
89.888.785.387.385.8
Dialects
Dialectal Arabic MMLU*
Accuracy
89.690.086.690.188.7
Dialect Identification (QADI)
Accuracy
71.162.352.860.560.8
Alyah*
Accuracy
87.682.479.381.077.9
Linguistics & Literature
Arabic MMLU with diacritics
Accuracy
88.288.484.688.887.4
FannOrFlop-MCQ
Accuracy
73.269.367.069.869.7
Arabic Linguistics^
Accuracy (mean of the 2 below)
94.894.088.095.093.3
Diacritization
Accuracy
91.992.088.493.292.3
Parsing
Accuracy
97.795.987.796.994.3
Translation
FLORES200+
LLM-as-a-judge
91.983.684.586.983.9
Assistant Capabilities
Instruction Following
Arabic IFBench^
Accuracy (prompt-level strict)
93.975.153.881.778.9
Arabic IFEval
Accuracy
77.774.570.179.279.2
Long Context
AA-LCR (Ar)
Accuracy
77.078.080.074.073.0
Factuality & Hallucination
HalluScore
Accuracy
70.968.863.966.871.2
SimpleQA Verified (Ar)
Aggregate (2/3 F-score + 1/3 non-hallucination rate)
31.235.832.820.422.9
F-score
8.553.026.528.632.9
Non-hallucination rate
76.71.545.33.93.0
Agentic
Terminal Bench 2.1
Accuracy (Avg@3)
73.851.365.286.153.9
GDPVal
Accuracy (@ 75% threshold)
73.244.570.976.860.0
MCPMark
Accuracy (Pass@1)
68.552.848.866.945.7
WildClawBench (English text-only subset)
Mean Score@3
50.446.656.459.034.2
Apex-Agents (420 text-only subset, loop agent)
Pass@1 (Accuracy)
18.615.323.836.49.0
GAIA v2 lit (Ar)
Accuracy (Pass@1)
50.345.544.250.440.0
Arabic File Processing
Accuracy (@ 75% threshold)
64.115.549.582.729.1
SWE-Bench Verified (Offline)
Accuracy (Pass@1)
70.271.875.680.867.8
BrowseComp
Discard-all@95k
72.678.683.579.844.4

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.

Quickstart

Serving

The commands below reflect our reference deployment: 16× NVIDIA H100 (80 GB) across two 8-GPU nodes.

vLLM (vllm>=0.30.0) · vLLM recipe:

vllm serve inception42/Jais-3-375B-A23B \
  --trust-remote-code \
  --dtype bfloat16 \
  --tensor-parallel-size 16 \
  --enable-expert-parallel \
  --reasoning-parser k2_horizon \
  --enable-auto-tool-choice \
  --tool-call-parser k2_horizon

Use an exact branch name from the inventory with vLLM's --revision option.

SGLang (sglang>=0.5.20) · SGLang recipe:

# Multi-node (2 nodes). Run the same command on every node with:
#   <node-rank> = 0 on the head node, 1 on the other node
#   <node0-ip>  = IP of the head node (reachable from all others)
sglang serve \
  --model-path inception42/Jais-3-375B-A23B \
  --tp 16 \
  --nnodes 2 \
  --node-rank <node-rank> \
  --dist-init-addr <node0-ip>:20000 \
  --ep 16 \
  --dtype bfloat16 \
  --attention-backend fa3 \
  --model-loader-extra-config '{"enable_multithread_load":false}' \
  --reasoning-parser k2_horizon \
  --host 0.0.0.0 \
  --port 30000

Running on H200. With 141 GB per GPU, the model fits on a single 8-GPU H200 node. In that case set tensor parallelism and expert parallelism to 8: --tensor-parallel-size 8 for vLLM, --tp 8 --ep 8 for SGLang.

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:30000/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="inception42/Jais-3-375B-A23B",
    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:30000/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="inception42/Jais-3-375B-A23B",
        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.

Transformers

Validated with Transformers 5.15.0, PyTorch 2.13.0, Safetensors 0.8.0.

from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "inception42/Jais-3-375B-A23B"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="auto", dtype="bfloat16", low_cpu_mem_usage=True, trust_remote_code=True
)
inputs = tokenizer("لدينا عددان مجموعهما 50 وحاصل ضربهما 600. ما هما العددان؟ اشرح خطوات الحل.", return_tensors="pt").to(model.device)
inputs.pop("token_type_ids", None)
outputs = model.generate(**inputs, max_new_tokens=32768, temperature=0.7, top_p=0.95, top_k=20, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Release Artifacts

The table below lists the release artifacts for Jais-3-375B-A23B.

Artifact Index

Artifact Link
Model card Hugging Face
Training logs W&B
Blog post Jais 3: Introducing Arabic-centric Agentic AI
Checkpoints Checkpoint inventory
Technical report Jais 3 Technical Report
Code repository Coming soon

Checkpoint Inventory

Model repository: inception42/Jais-3-375B-A23B

Branch names below refer to this repository.

Checkpoint Branch / repository
CPT Checkpoint CPT
Cooldown Checkpoint Cooldown
FTPO Checkpoint FTPO
SFT Checkpoint SFT

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. Serving. Use the validated SGLang recipe above: BF16, TP=16 and EP=16 across two 8× H100 nodes, FlashAttention-3, with multithreaded weight loading disabled. On a single 8× H200 node, use TP=8 and EP=8 instead.
  4. Parsers. Enable the k2_horizon reasoning parser for chat, and add the k2_horizon tool-call parser for agent use. Leave both off for plain completion-style generation.

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
1,760
Safetensors
Model size
379B params
Tensor type
BF16
·
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for inception42/Jais-3-375B-A23B

Quantizations
1 model

Collection including inception42/Jais-3-375B-A23B

Paper for inception42/Jais-3-375B-A23B