Tasks and inputs
Jais-3-375B-A23B is listed for Text generation, Translation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Jais-3-375B-A23B is the flagship of IFM's open Jais 3 family: a 375B MoE with 23B active and 512K context, leading on Arabic but heavy to serve.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Sixteen H100s. That is the reference deployment for Jais-3-375B-A23B, the flagship of the Jais 3 family, and it frames every other decision about this model. MBZUAI's Institute of Foundation Models and G42's Inception42 released it under Apache-2.0: a sparse mixture-of-experts model with 375B parameters, 23B active per token, and a 512K-token context.
The open question is not whether it is good at Arabic. It is whether it is good enough to justify the hardware over a smaller sibling or a rival frontier model.
| Option | Hardware | Parallelism |
|---|---|---|
| Reference | 16x H100 80 GB, two 8-GPU nodes | TP=16 with expert parallel |
| Single node | 8x H200 141 GB | TP=8, EP=8 |
An FP8 build (IFM/Jais-3-375B-A23B-FP8) quantises only the routed experts: static FP8 weights with one scale per 128x128 block and dynamic FP8 activations. Attention, shared experts, routers, the first three dense layers and the output head stay in BF16. IFM says the FP8 model tracks the BF16 original closely but publishes no numbers for the gap.
There is one hard constraint. The routed-expert intermediate size is 1,792, which does not split into whole 128-wide blocks at TP=4 or TP=8. Expert parallelism is therefore required, not optional. Metadata also lists 379B parameters rather than 375B.
IFM compared it with Qwen3.5 397B-A17B, MiniMax M3, Qwen3.8-Flash-Next and Nemotron 3 Ultra. The pattern is consistent.
Arabic language and culture: it leads on Belebele (95.6), IslamicQA, dialect identification, Arabic IFBench (93.9), FLORES200+ and parsing. Arabic MMLU (89.3) is a tie.
Open-ended Arabic generation: Qwen3.5 397B is ahead on most AraGen sub-scores.
Agentic coding: Qwen3.8-Flash-Next leads clearly. Terminal Bench 2.1 is 73.8 against 86.1; SWE-Bench Verified 70.2 against 80.8.
Factuality: an odd profile. Its SimpleQA non-hallucination rate is the highest in the table at 76.7, but its F-score is 8.5. It declines to answer a lot rather than guessing.
All publisher-reported.
Sovereign or regulated deployments that need the strongest open Arabic model on their own hardware, with a licence that permits commercial use. If you mostly need code agents, a smaller Qwen-family model will serve you better per GPU. If you mostly need Arabic chat at volume, try the sparse MoVA-36B-A4B sibling first; it activates a sixth as many parameters.
For the record, the flagship's repository dates to 29 September 2026, two weeks ahead of the reported 15 October family launch.
Run a two-week shadow test. Put the flagship behind your production prompts on a rented H200 node, log outputs next to your current model, and have Arabic-speaking reviewers grade a sample blind. Divide the quality gain by the GPU-hour bill. Requests need reasoning_effort: "high" and at least 32,768 output tokens, so include thinking time in your latency numbers. Other large open models are under text generation.
Compare the model on the work you need to complete.
Jais-3-375B-A23B is listed for Text generation, Translation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 524.29K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Tencent
Tencent Hy4 Flash is the cost-focused companion to Hy4, keeping vision and agent skills at a far lower token price. Availability is still unsettled.
Alibaba
Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.
Sarvam
Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.
IBM
Granite 4.1 8B LLMON Preview is IBM's model trained on block-structured LLMON prompts to resist indirect prompt injection in agents and RAG.
NVIDIA
Nemotron-Labs Sigma 8B is NVIDIA's open continuous diffusion language model, denoising 64-token blocks in parallel; English, GPU-only, 8K context.
Alibaba
Qwen4 Max Preview was Alibaba's five-day early-access build of its Qwen4 flagship, offered only through a Token Plan subscription before GA qwen4-max.