Jais 3-375B-A23B

Jais 3-375B-A23B is the flagship open-weights Arabic language model developed jointly by MBZUAI's Institute of Foundation Models (IFM) and Inception42. Released on October 15, 2026, it represents a major push toward sovereign, enterprise-grade AI infrastructure for the Middle East and North Africa. Built on the K2 Horizon foundation model family, the architecture uses a sparse Mixture of Experts (MoE) design that packs 375 billion total parameters while activating only 23 billion per token.

Unlike previous Arabic models that focused strictly on conversational fluency, Jais 3-375B-A23B is engineered explicitly for agentic workloads. It handles complex document parsing, multi-step tool use, and structured function-calling directly in Arabic script. It targets enterprise deployments requiring local data governance, serving as an on-premises replacement for proprietary frontier APIs in sectors like legal, finance, and regional government.

The model competes directly with large-scale frontier MoEs such as Qwen3.5-397B-A17B and MiniMax M3. By pairing its massive capacity with a native 512,000-token context window, it can ingest hundreds of pages of technical Arabic documentation, process regional code-switching, and maintain state across lengthy tool execution loops without context fragmentation.

Jais 3-375B-A23B uses a sparse Mixture of Experts (MoE) topology. Understanding how this architecture separates memory capacity from execution cost is critical when you plan to run Jais 3-375B-A23B locally:

• Total Parameters: 375 billion. All weights must reside in memory (system RAM or VRAM) to execute an inference pass.
• Active Parameters: 23 billion per token. During generation, the routing mechanism routes tokens through a subset of expert networks, capping dynamic compute requirements.
• Context Length: 512,000 tokens (512K).
• Base Architecture: K2 Horizon foundation family.
• Format Availability: Full precision (BF16), official FP8 checkpoint via Hugging Face (IFM/Jais-3-375B-A23B-FP8), and community quantized GGUF variants.

The practical implication of the Jais 3-375B-A23B MoE efficiency is twofold. First, your hardware pool must have enough raw capacity to hold the full 375B footprint. Quantized to 4-bit, the weights alone occupy roughly 210 GB to 230 GB. Second, because each forward pass calculates only 23 billion active parameters, compute throughput matches that of a mid-sized dense model. Once the weights are mapped to memory, the generation phase runs dramatically faster than a standard 375B dense architecture.

The 512K context window introduces substantial memory requirements for the Key-Value (KV) cache. If you leverage the full context window, you must account for tens of gigabytes of additional VRAM purely for attention state storage, especially without dynamic cache quantization (such as FP8 or FP4 KV cache schemes).

Jais 3-375B-A23B is trained on Modern Standard Arabic (MSA) alongside major regional dialects, including Emirati, Gulf, and North African Arabic. It also natively parses Arabizi and cross-lingual code-switching between Arabic, English, and French. Its core capabilities span:

The model is tuned for tool calling, external API execution, and structured data output (JSON/YAML). In production, it handles multi-hop operations such as querying enterprise SQL databases, parsing the retrieved rows into localized reports, and invoking external search engines for verification.

Given its 512K context buffer and specialized domain pre-training, the model processes large contractual archives, corporate filings, and regulatory frameworks. Benchmark results on Arabic Legal (HELM) and domain-specific evaluations demonstrate high factual accuracy when cross-referencing complex statutory texts.

On Arabic knowledge evaluations, the model scores highly on the IslamicQA suite (averaging over 90% accuracy) and sets a high baseline for Arabic MMLU (89.3%). It maintains strong contextual grounding in Islamic jurisprudence, regional history, and dialect-specific semantics, avoiding the common translation artifacts present in models that treat Arabic as a secondary translation layer.

Deploying a local AI model with 375B parameters in 2026 demands realistic hardware planning. While the 23B active parameter count keeps FLOP requirements modest, memory capacity remains the primary bottleneck.

The total memory footprint consists of base model weights plus KV cache overhead:

• FP16 / BF16 (16-bit): ~750 GB VRAM. Requires an 8x NVIDIA H100 (80GB) or 8x A100 (80GB) cluster using Tensor Parallelism (tp=8).
• FP8 (8-bit): ~390 GB to 420 GB VRAM. Requires at least 4x NVIDIA H200 (141GB) or a cluster of 8x 80GB enterprise GPUs.
• Q4_K_M (4-bit): ~215 GB to 240 GB total memory. Runs on dedicated multi-GPU workstations or unified memory workstations.
• KV Cache Overhead: Running short prompts (under 8K tokens) adds negligible overhead (4-8 GB). Utilizing the full 512K context can demand an additional 40 GB to 100+ GB of cache allocation depending on KV precision and attention head dimensioning.

If you are wondering how to run a 375B model on consumer GPU setups: a single consumer GPU cannot load this model. Even an RTX 4090 or RTX 5090 with 24GB or 32GB of VRAM cannot fit a fraction of the 375B weights.

To run Jais 3-375B-A23B locally on a workstation, consider these environments:

• Multi-GPU Workstation: 8x NVIDIA RTX 3090/4090 (24GB each) provides 192GB of total VRAM, which falls just short of a comfortable Q4 quantization footprint once runtime memory and context overhead are counted. A rig using 6x to 8x enterprise 48GB cards (such as RTX 6000 Ada or A6000) provides the 288GB to 384GB required to run 4-bit or 5-bit quantizations with substantial context windows.
• Apple Silicon (Unified Memory): An Apple Mac Studio or Mac Pro equipped with an M3/M4 Max or Ultra chip configured with 192GB of Unified Memory can run aggressively quantized variants (Q3_K_M), while a 256GB Unified Memory configuration comfortably hosts the recommended Q4_K_M quant with room for moderate context length.

• FP8 Checkpoint: Best balance for enterprise GPU racks (IFM/Jais-3-375B-A23B-FP8). It preserves full reasoning accuracy and dialect comprehension with almost zero degradation compared to BF16.
• Q4_K_M (Recommended for Local Workstations): The best quantization for Jais 3-375B-A23B on local hardware. It drops memory consumption down to ~220 GB while retaining high scores on Arabic MMLU and dialect instruction-following.
• Q2_K / Q3_K: Drops memory requirements below 180 GB, but degrades nuanced dialect handling, syntax cohesion, and tool-calling reliability.

For production GPU servers, run Jais using vLLM or TensorRT-LLM using tensor parallelism:

1vllm serve IFM/Jais-3-375B-A23B-FP8 \
2    --tensor-parallel-size 8 \
3    --max-model-len 32768 \
4    --trust-remote-code

For developers testing on single unified-memory host systems or heterogeneous environments, Ollama or raw llama.cpp provides the quickest way to get started using GGUF builds:

1ollama run jais-3:375b-q4

The Jais 3-375B-A23B tokens per second you achieve will scale directly with memory bandwidth rather than raw compute:

• 8x H100 (FP8 via vLLM): 35 to 55 tokens per second. The 23B active parameter path ensures extremely fast generation cycles once loaded.
• Apple Silicon M4 Max/Ultra (256GB Unified Memory, Q4_K_M via llama.cpp): 3 to 7 tokens per second. Memory bus bandwidth (typically 400 to 800 GB/s) serves as the primary speed gate.

Evaluating Jais 3-375B-A23B against other large-scale open MoE alternatives highlights distinct tradeoffs in regional adaptation versus broad-domain code and math:

• Architecture: Jais activates 23B out of 375B, whereas Qwen3.5 activates 17B out of 397B. Qwen has a slight edge in generation efficiency per token due to its smaller active slice.
• Arabic Native Reasoning: Jais 3 outperforms Qwen across Arabic-specific evaluations, particularly on regional cultural context (IslamicQA at 90.2% vs. 88.0%) and dialect comprehension (Egyptian, Emirati, and Maghrebi varieties).
• STEM and Code: Qwen3.5 holds a slight lead on purely English-centric coding tasks and international math benchmarks like GSM8K (93.6% vs. Jais's 86.8%).
• Selection Criteria: Choose Jais 3-375B-A23B if your pipeline processes Arabic-first text, legal discovery in MENA jurisdictions, or spoken dialect transcription transcripts. Choose Qwen3.5 if your primary workload is multilingual software engineering or dense math logic.

• Parameter Footprint: MiniMax M3 runs 428B total parameters with 23B active, presenting similar compute profiles but requiring even higher VRAM capacity (roughly 10% to 15% more memory).
• Context Handling: Both feature massive context designs, but Jais 3 natively incorporates regional Arabic vocabulary structures throughout its tokenizer, resulting in better token compression for Arabic source material. This lowers context memory pressure and speeds up ingestion throughput on Arabic legal, medical, and governmental corpuses.