Jais-3-7B-GGUF Arabic and English 7B Reasoning Model: 8GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | IFM/Jais-3-7B-GGUF |
| Family guide | Jais-3 guide (4 articles) |
| Publisher guide | IFM (MBZUAI): models and licenses |
| Published | 2026-10-13 |
| 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 “IFM" has released a model on Hugging Face under the name IFM/Jais-3-7B-GGUF. Please note that the editorial department has not been able to independently verify an official announcement for this release, and this is based solely on the description in the Hugging Face model card.
This repository is said to be a conversion of the base model inception42/Jais-3-7B into the GGUF format for use with llama.cpp. Jais-3-7B is a 7B-class decoder-only model intended for Arabic and English assistant and agent use cases, and the model card notes that it is positioned as a “medium dense" model within the Jais-3 family.
Specifications
Only the items listed in the model card of the base model (inception42/Jais-3-7B) are shown.
- Parameters: 7B
- Active/Effective parameters: 7B
- Architecture: Dense (
K2HorizonForCausalLM/ model_type:k2_horizon) - Context length: 512K
- Supported languages: Arabic, English
The GGUF version reportedly provides BF16, Q8_0, Q6_K, Q5_K_M, Q5_0, and Q4_K_M quantization and precision variants. Additionally, the model card warns that support for the “K2 Horizon" architecture used in the GGUF conversion was just recently merged into llama.cpp, and running it requires a relatively recent build.
Performance
Below is an excerpt from the comparison table published in the inception42/Jais-3-7B model card. Due to the number of columns, it focuses on Jais-3-7B and four comparable open-weight models that can be run locally (Gemma-4-E4B-it, Qwen3.5-9B, Falcon-H1-7B-Instruct, and Allam 7B). These are self-measured figures by the publisher (IFM).
→ Scroll horizontally to see all columns
| Metric | Jais-3-7B | Gemma-4-E4B-it | Qwen3.5-9B | Falcon-H1-7B-Instruct | Allam 7B |
|---|---|---|---|---|---|
| AraGen v3 (3C3H) 09-2025 | 55.0 | 48.8 | 55.7 | 28.0 | 44.9 |
| Arabic MMLU* Accuracy | 85.2 | 77.5 | 80.9 | 63.3 | 71.2 |
| Arabic GSM8K* Accuracy | 92.7 | 87.0 | 91.4 | 85.0 | 56.8 |
| Arabic IFEval Accuracy | 80.2 | 72.3 | 73.8 | 62.4 | 48.0 |
| AA-LCR (Ar) Accuracy | 65.0 | – | 68.0 | 9.0 | – |
| SimpleQA Verified (Ar) Aggregate | 25.7 | 24.4 | 8.5 | 3.8 | 4.2 |
| Terminal Bench 2.1 Accuracy (Avg@3) | 46.4 | 1.9 | 29.2 | 0.0 | 0.0 |
| GDPVal Accuracy (@ 75% threshold) | 60.5 | 5.5 | 19.1 | 0.0 | 0.0 |
In Arabic MMLU (four-choice general knowledge questions), Arabic GSM8K (step-by-step arithmetic word problem solving), and Arabic IFEval (format-based instruction following), Jais-3-7B scored the highest among these four models. It also ranks at the top for SimpleQA Verified (Ar)—which measures factual question-answering accuracy—with a score of 25.7, giving it an advantage in being less prone to fabricating information when it does not know the answer. In Terminal Bench 2.1 (which evaluates terminal operation task completion) and GDPVal (which measures business task achievement), Jais-3-7B significantly outperforms the other four models, highlighting its strength in agentic tasks. Meanwhile, in AA-LCR (Ar), which relates to long-form reading comprehension, Qwen3.5-9B (68.0) slightly edges out Jais-3-7B (65.0), and in AraGen v3, which evaluates open-ended generation, Qwen3.5-9B (55.7) also narrowly surpasses Jais-3-7B (55.0). The model card notes that these scores are based on the BF16 source checkpoint and that individual GGUF quantization variants have not been evaluated separately.
Strengths and Use Cases
Jais-3-7B is designed to provide advanced assistant and agent capabilities in both Arabic and English. Based on the original model card and performance data, it is considered particularly suited for deployment in the following areas:
Advanced Reasoning and Logical Thinking
The model is specifically designed for “Reasoning" and features the ability to generate a thought process when solving complex problems. The publisher strongly recommends setting reasoning_effort="high" during use. This reportedly enables detailed, step-by-step responses for mathematical word problems and logic puzzles. As shown in the performance section table, it records exceptionally high scores in Arabic GSM8K (arithmetic) and GATmath (mathematics), indicating strength in tasks requiring logical precision.
Agent Tasks and Tool Use
Another major feature is its enhanced “Tool Calling" capability, allowing it to invoke and execute external tools. It supports tool calls in multiple formats such as XML and JSON, and reported benchmark results show it significantly outperforming other models of the same scale in agentic tasks like terminal operations. It is well-suited for developing applications that integrate with external functions, such as fetching the time for a specific city via an API.
Ultra-Long Context Processing
Equipped with an exceptionally large 512K context window, it enables the ingestion of massive documents and inference while retaining long conversation histories. It also shows solid performance in AA-LCR (Ar), a benchmark for long-form reading comprehension, making it promising for use cases handling large-scale knowledge bases.
Operational Limitations
On the other hand, the model card explicitly lists the following use cases as “Out of scope." It is not recommended for making critical decisions related to medical, legal, financial, or safety fields without human supervision, nor for applications requiring reliability in languages other than Arabic and English. Furthermore, because it may generate inaccurate content or plausible falsehoods (hallucinations), pre-deployment evaluation and the establishment of safety measures are necessary.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 9.0B parameters (taken from the base model inception42/Jais-3-7B)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 8GB (RTX 4060 / 3060 Ti, etc.) | Q5_K_M | 6.0GB | 7.2GB |
| 12GB (RTX 4070 / 3060 12GB, etc.) | Q8_0 | 8.9GB | 10.7GB |
| 24GB (RTX 4090 / 3090, etc.) | BF16 | 16.8GB | 20.1GB |
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.
Compression: the Q5_K_M build measures 5.73 bits per weight — about 36% the size of the original 16-bit weights, calculated by this site from the actual file sizes.
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 4–15B 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 |
|---|---|---|---|---|
| microsoft/UserRepGen-4B | 4.5B | 4GB | mit | UserRepGen-4B Text Generation Model: 4GB+ VRAM, GGUF Builds (2026-11-17) |
| IFM/Jais-3-3.7B-GGUF | 5.1B | 4GB | apache-2.0 | Jais-3-3.7B-GGUF 512K Context Language Model: 4GB+ VRAM (2026-11-17) |
| ibm-granite/granite-switch-4.1-8b | 8.8B | 24GB | apache-2.0 | IBM Releases Granite Switch 4.1: Modular LLMs with aLoRA and Speech (2026-11-17) |
| ggml-org/NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF | 8.0B | 12GB | apache-2.0 | NVIDIA-Nemotron-3.5-Bolt-E4B-GGUF Text Generation Model: 12GB+ VRAM (2026-11-17) |
| ByteDance-Seed/Cola2 | 8.8B | 48GB | — | ByteDance-Seed Releases Cola2 Language Model (2026-11-17) |
How to Get It
This model is available on the Hugging Face repository IFM/Jais-3-7B-GGUF. It is licensed under Apache 2.0, with no gating or terms-of-service agreement required for access, allowing anyone to download it.
It is distributed in the GGUF format with the following precision and quantization variants available:
– BF16
– Q8_0
– Q6_K
– Q5_K_M
– Q5_0
– Q4_K_M
The supported engine is llama.cpp. However, because support for the “K2 Horizon" architecture used by this model was just recently merged into llama.cpp, it is recommended to use a binary built from the latest source code.
An example command for launching it using llama-server is as follows:
./build/bin/llama-server -hf IFM/Jais-3-7B-GGUF:Q8_0 --jinja -ngl 99 -c 32768
When running, appending the --jinja flag enables the chat template properly configured to handle the reasoning process and tool calls. Additionally, the recommended settings are to specify reasoning_effort="high" during inference, along with sampling parameters of temperature=0.2 (1.0 in English and 0.7 in Arabic during thinking mode), top_p=0.95, and top_k=20. When performing tool calls, enabling --tool-call-parser k2_horizon is required to correctly parse formats like XML.
Related Articles
- Jais-3-3.7B-GGUF 512K Context Language Model: 4GB+ VRAM
- Jais-3-0.9B-GGUF Bilingual Dense Model: 4GB+ VRAM
- Jais-3-32B-GGUF Arabic and English 32B Dense Model: 24GB+ VRAM
- Jais-3-MoVA-36B-A4B-GGUF Text Generation Model: 32GB+ VRAM
What to Read Next
- Find models by VRAM (This model runs from the 8GB tier) → Other models that run on a 8GB GPU
- Explore the same model family → Jais-3 family overview (4 articles)
- Engines that run this model → llama.cpp / Ollama
- What Q5_K_M, Q8_0 mean and where to get this model → GGUF format guide and models / Quantization and model-format glossary
- Learn about the publisher → IFM (MBZUAI): models, licenses and articles
Sources
This article contains unverified information. We will append an update note once it is confirmed by a primary source.

