Jais-3-MoVA-36B-A4B-GGUF Text Generation Model: 32GB+ VRAM

At a Glance
| Item | Value |
|---|---|
| Repository | IFM/Jais-3-MoVA-36B-A4B-GGUF |
| Publisher guide | IFM (MBZUAI): models and licenses |
| Published | 2026-10-12 |
| License | apache-2.0 |
| Formats | GGUF |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
IFM has released GGUF format models for “Jais-3-MoVA-36B-A4B", a large language model for Arabic and English developed by inception42.
Jais-3-MoVA-36B-A4B adopts a Mixture-of-Experts (MoE) architecture featuring Mixture-of-Values attention (MoVA), operating with 4B active parameters per token out of a total of 36B parameters. It also supports an extensive context length of 512K tokens.
In this repository, various quantization variants from BF16 down to Q4_K_M are provided in GGUF format compatible with llama.cpp, allowing selection to match local hardware configurations.
Specifications
- Total parameters: 36B
- Activated parameters: 4B (per token)
- Architecture: MoE (Mixture-of-Experts with Mixture-of-Values attention / K2HorizonForCausalLM)
- Context length: 512K
Performance
Performance data listed in the model card are measurement results from the publisher on the original BF16 checkpoint, rather than unique measured values of the quantized versions. The model card notes that individual evaluations for the quantization variants have not yet been conducted.
Comparison tables with open-weight models of a similar scale published in the model card are shown below (some columns and rows are based on the model card descriptions).
→ Scroll horizontally to see all columns
| Domain / Benchmark Item | Jais-3-MoVA-36B-A4B | Falcon-H1-34B-Instruct | Gemma-4-31B-it | Muse-Glimmer-30B | Qwen3.8-27B | Fanar-2-27B |
|---|---|---|---|---|---|---|
| Architecture Information | ||||||
| Total parameters (# Params) | 36B | 34B | 31B | 30B | 27B | 27B |
| Activated parameters (# Activated params) | 4B | 34B | 31B | 30B | 27B | 27B |
| Architecture | MoE | Dense | Dense | Dense | Dense | Dense |
| Open-ended Generation | ||||||
| AraGen v3 (3C3H) 09-2025 | 68.6 | 45.0 | 73.4 | 75.6 | 66.7 | 55.9 |
| Correctness | 79.7 | 53.2 | 85.4 | 87.7 | 77.1 | 65.7 |
| Completeness | 78.3 | 51.0 | 84.9 | 87.0 | 76.2 | 63.5 |
| Conciseness | 24.4 | 19.7 | 21.1 | 23.0 | 23.9 | 21.6 |
| Helpfulness | 72.6 | 45.4 | 79.7 | 81.2 | 71.5 | 59.0 |
| Honesty | 77.9 | 49.0 | 84.4 | 86.7 | 74.7 | 61.3 |
| Harmlessness | 78.8 | 51.4 | 84.9 | 87.7 | 76.5 | 64.4 |
| Knowledge & Reasoning | ||||||
| Arabic MMLU* Accuracy | 87.7 | 73.7 | 89.5 | 87.0 | 85.8 | 66.3 |
| Belebele Accuracy | 94.4 | 86.6 | 95.9 | 94.3 | 94.4 | 61.8 |
| AraSTEM* Accuracy | 94.7 | 85.5 | 96.3 | 94.7 | 95.3 | 42.4 |
| Arabic GSM8K* Accuracy | 92.2 | 90.2 | 93.7 | 95.2 | 93.2 | 87.7 |
| GATmath* Accuracy | 97.5 | 63.4 | 98.6 | 98.5 | 98.4 | 87.8 |
| MedAraBench* Accuracy | 85.4 | 60.7 | 88.8 | 86.1 | 87.3 | 50.9 |
| Arabic Legal (HELM)* Accuracy | 69.3 | 63.1 | 71.1 | 73.1 | 70.5 | 45.7 |
| Arabic Language & Culture | ||||||
| IslamicQA Accuracy | 89.7 | 71.1 | 89.7 | 83.7 | 77.1 | 52.0 |
| QIAS-1* Accuracy | 95.7 | 56.2 | 93.2 | 87.4 | 81.6 | 31.4 |
| Arab Culture Accuracy | 87.5 | 76.5 | 90.2 | 84.6 | 81.0 | 41.4 |
| Dialectal Arabic MMLU* Accuracy | 87.3 | 71.5 | 89.4 | 87.0 | 87.0 | 28.2 |
| Arabic Linguistics^ Accuracy | 94.0 | 57.4 | 96.7 | 91.8 | 89.7 | 39.1 |
| FLORES200+ LLM-as-a-judge | 90.8 | 70.6 | 92.8 | 84.8 | 81.8 | 71.2 |
| Assistant Capabilities | ||||||
| Arabic IFBench^ Accuracy | 92.0 | 32.2 | 80.8 | 84.2 | 81.3 | 40.0 |
| Arabic IFEval Accuracy | 80.0 | 66.8 | 78.7 | 79.5 | 77.0 | 49.8 |
| AA-LCR (Ar) Accuracy | 73.0 | 14.0 | 66.0 | – | 79.0 | – |
| Factuality & Hallucination | ||||||
| HalluScore Accuracy | 66.0 | 42.4 | 69.7 | 69.1 | 54.6 | 41.7 |
| SimpleQA Verified (Ar) Non-hallucination rate | 83.1 | 9.4 | 10.5 | 3.7 | 5.4 | 12.2 |
| SimpleQA Verified (Ar) Aggregate | 31.5 | 7.5 | 10.6 | 13.3 | 8.0 | 11.0 |
| SimpleQA Verified (Ar) F-score | 5.7 | 6.6 | 10.7 | 18.2 | 9.2 | 10.4 |
| Agentic | ||||||
| Terminal Bench 2.1 Accuracy (Avg@3) | 67.3 | 1.1 | 43.5 | 51.7 | 79.8 | 1.1 |
| GDPVal Accuracy (@ 75% threshold) | 67.7 | 0.0 | 30.5 | 48.2 | 78.2 | 0.0 |
From these comparison results, the following points regarding the strengths and weaknesses of the original model can be drawn:
In “Arabic IFEval", which measures instruction-following performance, it recorded 80.0, demonstrating top-tier levels surpassing Gemma-4-31B-it (78.7) and Qwen3.8-27B (77.0). Furthermore, it reaches 92.0 in “Arabic IFBench", showing a significant advantage over the competing Dense models.
In addition, the non-hallucination rate for “SimpleQA Verified (Ar)", which indicates the degree to which fabricated stories are prevented against factual queries, is exceptionally high at 83.1, leaving a prominent result compared to other models which all remain around 10%. In tasks specific to the Arabic language, tendencies toward strong cultural and regional knowledge are exhibited, scoring 89.7 in IslamicQA and 95.7 in QIAS-1.
On the other hand, in “Arabic GSM8K" for solving elementary school-level math problems, it stands at 92.2, slightly trailing Muse-Glimmer-30B (95.2) and Gemma-4-31B-it (93.7). Furthermore, it recorded 67.3 in “Terminal Bench 2.1", which evaluates task execution capability for terminal operations; while it significantly outperforms some models like Falcon-H1-34B-Instruct (1.1), it falls far behind Qwen3.8-27B (79.8). Aspects where it falls short of competing powerful Dense models can be seen in agent tasks and certain mathematical domains.
Strengths and Use Cases
Jais-3-MoVA-36B-A4B is designed for assistant and agent use cases targeting Arabic and English. It is intended for use in reasoning tasks, complex workflows involving tool calling, as well as various research and experiments.
Main strengths and features seen from the original model’s design and performance evaluations are as follows:
-
Advanced reasoning and assistant functions in Arabic and English
It demonstrates accuracy exceeding competing models of a similar scale in instruction following (Arabic IFBench and Arabic IFEval), and responses that accurately follow a user’s detailed constraints and format specifications can be expected. It also features a reasoning capability that separates and outputs the thought process, making it suitable for solving complex problems requiring step-by-step thinking. -
Flexible tool calling
It supports the tool calling functionality of the K2 Horizon architecture, allowing agent processing such as external function execution and API integration. In addition to the default XML format (xml), call formats correspond tojsonandxml_typed(which retains type information), enabling switching according to the usage environment or framework. -
Regional, cultural, and linguistic knowledge of the Arabic-speaking sphere
It records extremely high scores in tasks specialized for the Arabic-speaking sphere, such as Islam-related knowledge (IslamicQA), Arab regional cultural understanding (Arab Culture), and regional dialects (Dialectal Arabic MMLU). It holds strengths in text generation reflecting regional-specific contexts and cultural nuances that cannot be fully covered by multilingual models. -
High factuality and hallucination suppression
It achieves an outstanding non-hallucination rate (83.1) in “SimpleQA Verified (Ar)", which is an index for avoiding fabricated stories against factual queries. The risk of carelessly generating answers contrary to facts regarding uncertain information is suppressed.
On the other hand, the official model card clearly specifies the following limitations and out-of-scope uses:
- Use in unsupported languages: Reliable operation in languages other than Arabic and English is not supported.
- High-risk decision-making: Relying on the model for critical judgments in medicine, law, finance, safety management, etc., without appropriate human oversight is not intended.
- Harmful and illegal uses: The generation of discriminatory, extremist, or intentionally misleading content, as well as use promoting illegal acts, is prohibited.
- Output accuracy and bias: Since incorrect information and biases resulting from training data may be included, prior verification and the application of safeguards are recommended for practical deployment.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 37.4B parameters (taken from the base model inception42/Jais-3-MoVA-36B-A4B)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 32GB (RTX 5090, etc.) | Q5_K_M | 24.6GB | 29.5GB |
| 48GB (RTX 6000 Ada / A6000, etc.) | Q8_0 | 37.1GB | 44.5GB |
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 5.64 bits per weight — about 35% 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
This model is available from the Hugging Face IFM/Jais-3-MoVA-36B-A4B-GGUF repository. The Apache-2.0 license is applied, and it can be downloaded and used without requiring prior agreement to terms of use (gated application).
The distribution format is GGUF, and the following quantization variants are provided:
– BF16
– Q8_0
– Q6_K
– Q5_K_M
– Q5_0
– Q4_K_M
Running with llama.cpp
Support for the K2 Horizon architecture adopted by this model has been merged into llama.cpp via pull request ggml-org/llama.cpp#29535. Execution requires the latest llama.cpp build or a binary built from the master branch.
Example build and inference server (llama-server) startup commands are as follows:
cmake -B build -DGGML_CUDA=ON
cmake --build build -j --target llama-server
./build/bin/llama-server -hf IFM/Jais-3-MoVA-36B-A4B-GGUF:Q8_0 --jinja -ngl 99 -c 32768 -np 1
Passing the --jinja option at server startup enables the chat template embedded in the model. This automatically parses the control of reasoning depth (reasoning_effort) and tool calling in K2 Horizon format, returning the thought content (reasoning_content) and final answer (content) separately.
Recommended Sampling Parameters
Setting parameters recommended by the publisher are as follows. All reported scores are measured with the reasoning_effort="high" setting.
- Reasoning depth (reasoning_effort): Specify
highin principle ({"reasoning_effort": "high"}) - Temperature (temperature):
- Default:
0.2– Thinking mode (English):1.0– Thinking mode (Arabic):0.7 - Top-P / Top-K:
top_p=0.95,top_k=20 - Penalty related:
min_p=0.0,presence_penalty=0.0,repetition_penalty=1.0 - Maximum output tokens: 32,768 tokens or more
Other Distribution Formats
In IFM’s Hugging Face collection, formats other than GGUF are also published:
– FP8 quantized version: IFM/Jais-3-MoVA-36B-A4B-FP8
– For Apple Silicon (MLX):
– IFM/Jais-3-MoVA-36B-A4B-MLX-4bit – IFM/Jais-3-MoVA-36B-A4B-MLX-8bit – IFM/Jais-3-MoVA-36B-A4B-MLX-bf16
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-3-32B-GGUF Arabic and English 32B Dense Model: 24GB+ VRAM
What to Read Next
- Find models by VRAM (This model needs at least 32GB) → VRAM quick reference
- 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
- Other models for the same task → Other text generation models

