Jais-3-0.9B-GGUF Bilingual Dense Model: 4GB+ VRAM

Jais-3-0.9B-GGUF Bilingual Dense Model: 4GB+ VRAM

At a Glance

Item Value
Repository IFM/Jais-3-0.9B-GGUF
Family guide Jais-3 guide (4 articles)
Publisher guide IFM (MBZUAI): models and licenses
Published 2026-10-12
License apache-2.0
Formats GGUF
Source type Unverified (not confirmed by a primary source)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

It is reported that a GGUF quantized version of “Jais-3-0.9B", the smallest model in the “Jais-3" family specializing in Arabic-English bilingual performance, has been released. However, note that this information has not been officially confirmed at this time and is based on unverified information from repository availability. Despite its extremely lightweight size of 0.9B (approx. 900 million) parameters, this model features a long 128K context window and is designed for reasoning and agent use cases.

The original model, “Jais-3-0.9B", is a decoder-only Dense model developed with the goal of realizing agent AI focused on Arabic. Tags suggest that knowledge distillation techniques were used, and it is expected to have advanced reasoning capabilities despite its small scale. With this GGUF release, it is reported that running it in general PC environments using lightweight inference engines like llama.cpp becomes possible.

Specifications

  • Parameters: 0.9B
  • Architecture: Dense (K2 Horizon)
  • Context Length: 128K

Performance

Regarding the performance of this model, benchmark results for the original “Jais-3-0.9B" model provided by the publisher are introduced below. While slight degradation may occur in the quantized version compared to these numbers, they serve as an important indicator for understanding the model’s characteristics. Models of similar to slightly larger sizes, such as Qwen3.5-2B and Falcon-H1-1.5B, are listed for comparison.

→ Scroll horizontally to see all columns

Metric Jais-3-0.9B Qwen3.5-2B Falcon-H1-1.5B-Deep-Instruct Hala 1.2B tiny-aya-global
Open-ended Generation
AraGen v3 (3C3H) 32.7 15.3 12.2 25.0 21.4
Knowledge & Reasoning
Arabic MMLU* Accuracy 70.4 63.1 48.7 45.0 55.9
STEM & Math
AraSTEM* Accuracy 81.6 83.9 62.5 48.3 58.5
Arabic GSM8K* Accuracy 80.9 75.6 65.4 57.3 59.6
Professional Domains
MedAraBench* Accuracy 48.1 54.3 32.3 32.3 39.5
Arabic Language & Culture
IslamicQA Accuracy (mean) 66.2 43.0 48.2 35.9 44.2
Arab Culture Accuracy (mean) 73.3 54.8 54.4 46.9 61.5
Assistant Capabilities
Arabic IFBench^ Accuracy 91.1 33.1 16.2 14.6 21.5
Arabic IFEval Accuracy 71.3 62.6 49.3 44.1 58.2
Factuality & Hallucination
Non-hallucination rate 79.3 18.2 3.8 3.2 3.2

According to measurements by the publisher, this model shows performance that overwhelms same-sized models across many Arabic-related tasks. Particularly noteworthy is the instruction-following accuracy in “Arabic IFBench", where the figure of 91.1% significantly distances Qwen3.5-2B (33.1%), the highest among the comparison targets. Additionally, the “Non-hallucination rate" in factuality metrics is extremely high at 79.3%, suggesting a strong ability not to fabricate answers when information is unknown.

In general knowledge tests like “Arabic MMLU" and math word problems like “Arabic GSM8K", it also records scores equal to or higher than 2B-class models with more parameters than itself. On the other hand, in STEM fields (AraSTEM) and medical fields (MedAraBench), it yields to Qwen3.5-2B, showing that larger models have the advantage in specialized science and technology fields. However, it consistently maintains a very high advantage in Arabic cultural understanding and linguistic tasks.

These scores back up the model’s exceptionally high suitability for assistants targeting Arabic-speaking users and agent tasks requiring adherence to specific formats. Achieving both such high instruction following and factuality at a size of 0.9B is a major advantage for local deployment.

Strengths and Use Cases

According to the original model’s design and the publisher’s materials, this model is developed for assistant and agent use cases supporting Arabic and English. It is especially envisioned for tasks involving reasoning and tool use, as well as application to research and experimentation.

Specifically, it is stated to be suited for the following fields and uses:

  • Tasks requiring advanced reasoning: Equipped with reasoning capabilities that explicitly output the thought process, enabling step-by-step logical responses.
  • Agent and tool integration: Compatible with tool-calling templates based on the K2 Horizon architecture, making it suitable for autonomous processing linked with external tools.
  • Arabic-centric dialogue and knowledge processing: Demonstrates high aptitude in Arabic-specific cultural and linguistic tasks, such as regional understanding of Islamic culture and the Arab world, dialect identification, Arabic grammar processing, and vowel diacritics (tashkeel).
  • Strict instruction following: High capability to generate responses without breaking instructed constraints or formats, suitable for workflows requiring structured output.
  • Long-context processing: Supports a long context window of 128K, enabling reference to lengthy documents and multi-turn conversation history.

On the other hand, the model card explicitly states the following out-of-scope uses and limitations:

  • Out-of-scope uses: General actions violating laws and regulations, generation of discriminatory, harmful, or extreme content or intentional false information, exposure or inference of personal or confidential information, and standalone use for high-stakes decisions involving life or safety such as medical, legal, financial, or safety management.
  • Language limitations: Reliable performance is not guaranteed for languages other than Arabic and English, and they are unsupported.
  • Output limits and bias: Because of the possibility of outputting inaccurate information or ungrounded content (hallucinations) and biases stemming from training data, validation and the introduction of safety measures are recommended before deployment to real environments.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 1.1B parameters (taken from the base model inception42/Jais-3-0.9B)

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) BF16 2.0GB 2.4GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp as-is.

It is distributed in GGUF, so no conversion is needed.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Recent Models in the Same Size Class

Models with up to 4B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
openbmb/MiniCPM-V-4.7-35B-A3B-DSpark-gguf 389M 4GB apache-2.0 MiniCPM-V-4.7-35B-A3B-DSpark-gguf Draft Model for Speedup: 4GB+ VRAM (2026-11-17)
openbmb/MiniCPM-V-4.7-1B-DSpark-gguf 214M 4GB apache-2.0 MiniCPM-V-4.7-1B-DSpark-gguf Text Generation Model: 4GB+ VRAM (2026-11-17)
nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16 4.0B 4GB openmdw-1.1 NVIDIA Releases Dory1.0-4B: A Recurrent 4B LLM with Dynamic Loops (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-DSpark 328M 4GB lfm1.0 LFM2.5-24B-A2B-DSpark Speculative Decoding Draft Model: 4GB+ VRAM (2026-11-17)
LiquidAI/LFM2.5-24B-A2B-DSpark-GGUF 328M 4GB lfm1.0 LFM2.5-24B-A2B-DSpark-GGUF Text Generation Model: 4GB+ VRAM (2026-11-17)

How to Get It

This model is distributed in GGUF format, and it is reported that multiple quantization variations such as BF16, Q8_0, Q6_K, Q5_K_M, Q5_0, and Q4_K_M are provided. The Apache-2.0 license applies, and since Hugging Face terms-of-use agreement (gating) is not set, it can be downloaded and used directly.

llama.cpp can be used as the inference engine. Because support code for the K2 Horizon architecture adopted by this model has already been merged into the llama.cpp repository, using the latest build or a binary built from the master branch is recommended.

At startup, to properly interpret thought processes and tool calls, the --jinja option must be provided to enable the built-in chat template. An example command for building and starting llama-server is as follows:

cmake -B build -DGGML_CUDA=ON
cmake --build build -j --target llama-server
./build/bin/llama-server -hf IFM/Jais-3-0.9B-GGUF:Q8_0 --jinja -ngl 99 -c 32768

For variations, specify tags like :BF16, :Q8_0, :Q6_K, :Q5_K_M, :Q5_0, or :Q4_K_M, or specify downloaded local files with the -m option.

After starting the local server, inference can be executed via an OpenAI-compatible API. A Python invocation example is as follows:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="Jais-3-0.9B",
    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)

As best practices recommended by the publisher, the following parameters are suggested:

  • Reasoning effort: Since benchmark results are all measured at a high thinking level, passing {"reasoning_effort": "high"} via chat_template_kwargs per request is recommended.
  • Sampling settings: Normally, temperature=0.2, top_p=0.95, and top_k=20 are recommended. When enabling thinking mode, setting temperature=1.0 for English and temperature=0.7 for Arabic is recommended. Maintaining min_p=0.0, presence_penalty=0.0, and repetition_penalty=1.0 is also guided.
  • Parser settings: By starting with --jinja enabled, thought processes and tool calls are automatically parsed on the llama.cpp side.

Note that for Apple Silicon environments, MLX-format variants (4bit, 8bit, bf16) are also reported to be separately released through a collection.

Related Articles

What to Read Next

Sources

This article contains unverified information. We will append an update note once it is confirmed by a primary source.