Jais-3-32B-GGUF Arabic and English 32B Dense Model: 24GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | IFM/Jais-3-32B-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, an AI development company in the Arabic-speaking world, has released “IFM/Jais-3-32B-GGUF", a repository converting the large dense model of the Jais-3 family “Jais-3-32B" into GGUF format on Hugging Face. However, official announcements regarding this matter have not yet been confirmed, and the contents of this article are unverified breaking news based on the model card information on Hugging Face.
According to the model card, the original model Jais-3-32B was created assuming assistant and agent use cases supporting both Arabic and English, as well as reasoning and tool-use tasks. This GGUF version enables running it in local environments using llama.cpp.
Specifications
- Number of Parameters: 32B (32B class)
- Architecture: Dense (decoder-only), model type is
k2_horizon(K2 Horizon architecture) - Context Length: 512K
Performance
Below is an excerpt from the comparison table published in the Jais-3-32B model card. Due to the large number of columns, it focuses on Jais-3-32B and four comparison models (Falcon-H1-34B-Instruct, Gemma-4-31B-it, Qwen3.8-27B, Fanar-2-27B), displaying only representative rows. Note that these figures were measured by the publishers of the original model Jais-3-32B (BF16 checkpoint), and the model card explicitly states that the GGUF quantized versions have not been evaluated separately.
→ Scroll horizontally to see all columns
| Jais-3-32B | Falcon-H1-34B-Instruct | Gemma-4-31B-it | Qwen3.8-27B | Fanar-2-27B | |
|---|---|---|---|---|---|
| # Params | 32B | 34B | 31B | 27B | 27B |
| Architecture | Dense | Dense | Dense | Dense | Dense |
| Arabic MMLU* Accuracy | 88.6 | 73.7 | 89.5 | 85.8 | 66.3 |
| Arabic GSM8K* Accuracy | 94.3 | 90.2 | 93.7 | 93.2 | 87.7 |
| Arabic IFEval Accuracy | 80.0 | 66.8 | 78.7 | 77.0 | 49.8 |
| SimpleQA Verified (Ar) Aggregate | 28.1 | 7.5 | 10.6 | 8.0 | 11.0 |
| Terminal Bench 2.1 Accuracy (Avg@3) | 63.7 | 1.1 | 43.5 | 79.8 | 1.1 |
From this table, Jais-3-32B records the highest value among the compared models in SimpleQA Verified (Ar) at 28.1. SimpleQA measures accuracy on factual questions with a single definite answer and looks at whether the model fills in unknowns with fabrications (factuality), which suggests that Jais-3-32B tends to make fewer unfounded answers than other models on Arabic factual questions. In Arabic IFEval, which measures adherence to instruction formats, it also tops the comparison group at 80.0, indicating high instruction-following accuracy. On the other hand, in Terminal Bench 2.1, which measures whether tasks can be completed by executing commands in a terminal, it scores 63.7, falling short of Qwen3.8-27B’s 79.8, meaning it is not necessarily the strongest in agent tasks involving terminal operations. Furthermore, in Arabic MMLU, which measures general knowledge, it slightly trails Gemma-4-31B-it’s 89.5 with 88.6, showing a narrow deficit in knowledge-based tasks.
Strengths and Use Cases
According to the model card of the original model Jais-3-32B, it is designed for assistant and agent use cases supporting both Arabic and English, reasoning tasks, and tool-use tasks. The tags also include reasoning and knowledge-distillation, suggesting it is intended for use cases that yield answers through thought processes rather than simple one-shot responses.
In fact, the model card recommends setting reasoning_effort to “high", with reasoning results returned separately as reasoning_content and final answers as content. For tool calling, it supports multiple formats such as json, xml, and xml_typed, making it suitable for agent-like workflows where tools are called while proceeding with tasks.
From the performance table mentioned earlier, it achieves top values among the comparison models in Arabic factual question accuracy (SimpleQA Verified (Ar)) and instruction following (Arabic IFEval), making it suitable for applications requiring accurate responses and instruction compliance in Arabic. On the other hand, it may underperform compared to other models in agent tasks involving terminal operations (Terminal Bench 2.1), so excessive expectations should probably not be placed on it for coding and system operation agent use cases.
Additionally, the model card explicitly states as “out of scope" reliance on languages other than Arabic and English, and unsupervised use in high-risk decision-making such as medical, legal, and financial domains. Human supervision is stated to be a prerequisite for these applications.
A long context length of 512K is also listed in the specifications, suggesting potential applicability for long-document processing and agent operations maintaining long conversation histories.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 34.8B parameters (taken from the base model inception42/Jais-3-32B)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 24GB (RTX 4090 / 3090, etc.) | Q4_K_M | 19.6GB | 23.6GB |
| 32GB (RTX 5090, etc.) | Q6_K | 26.6GB | 31.9GB |
| 48GB (RTX 6000 Ada / A6000, etc.) | Q8_0 | 34.4GB | 41.3GB |
| 80GB class (A100 / H100) | BF16 | 64.8GB | 77.8GB |
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 Q4_K_M build measures 4.84 bits per weight — about 30% 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 15–40B 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 |
|---|---|---|---|---|
| LiquidAI/LFM2.5-24B-A2B-Base | 24.0B | 80GB | lfm1.0 | LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-FP8 | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-FP8 Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-bf16 | 24.0B | 80GB | lfm1.0 | LFM2.5-24B-A2B-MLX-bf16 Reasoning Model for Apple Silicon: 80GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B-MLX-8bit | 24.0B | 32GB | lfm1.0 | LFM2.5-24B-A2B-MLX-8bit Text Generation Model: 32GB+ VRAM (2026-11-17) |
| LiquidAI/LFM2.5-24B-A2B | 24.0B | 12GB | lfm1.0 | LFM2.5-24B-A2B Agent-Optimized MoE Reasoning Model: 12GB+ VRAM (2026-11-17) |
How to Get It
The newly released repository is “IFM/Jais-3-32B-GGUF", which converts the original model Jais-3-32B into GGUF format for llama.cpp. The distributed quantization variations are 6 types: BF16, Q8_0, Q6_K, Q5_K_M, Q5_0, and Q4_K_M.
When running with llama.cpp, the model card provides the following command example:
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
You can launch it by specifying any of the tags like :BF16, :Q8_0, :Q6_K, :Q5_K_M, :Q5_0, :Q4_K_M, or by specifying a local file with -m.
According to warnings in the model card, support for the “K2 Horizon" architecture adopted by this model has already been integrated into llama.cpp (ggml-org/llama.cpp#29535), but it is noted that you need to use the latest build, build it yourself from master, or wait for the next official release. Also, note that the --jinja flag must be passed when starting llama-server to correctly parse reasoning content and tool calls.
The license is apache-2.0, and agreement to the terms of service on Hugging Face does not appear to be required.
Related Articles
- Jais-3-0.9B-GGUF Bilingual Dense Model: 4GB+ VRAM
- Jais-3-3.7B-GGUF 512K Context Language Model: 4GB+ VRAM
- Jais-3-7B-GGUF Arabic and English 7B Reasoning Model: 8GB+ 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 24GB tier) → Other models that run on a 24GB GPU
- Explore the same model family → Jais-3 family overview (4 articles)
- Engines that run this model → llama.cpp / Ollama
- What Q4_K_M, Q6_K, 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.

