What it does
Jais 3 is the third generation of the Jais open-weight language models, published by Inception (inception42) and IFM on Hugging Face under the Apache 2.0 licence. The family runs from a 0.9B dense model to the 375B-parameter Mixture-of-Experts flagship, all with a 512K context window, and is built for Arabic and English assistant and agentic applications with reasoning and tool use.
Official source 1 ↗ · Official source 2 ↗Features
- Model family from 0.9B to 375B. Ten checkpoints in the collection: 0.9B, 3.7B, 7B and 32B dense models (each also in an Uno variant), the sparse MoVA-36B-A4B and the 375B-A23B flagship. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
- Arabic and English. The model cards list Arabic and English as the supported languages and target Arabic and English assistant applications. Official source 1 ↗
- 512K context window. Every model card in the family describes a 512K-token context window. Official source 1 ↗ · Official source 2 ↗ · Official source 3 ↗
- Reasoning with adjustable effort. Reasoning depth is chosen per request; the thinking is returned separately from the answer. IFM recommends high reasoning effort, which all reported results use. Official source 1 ↗
- Tool calling. Supports tool calls in JSON, XML or typed-XML format for agent use, parsed by the k2_horizon tool-call parser. Official source 1 ↗
- Efficient sparse models. The MoE models only run part of their weights per token: 23B of 375B for the flagship and 4B of 36B for MoVA-36B-A4B, which uses Mixture-of-Values attention. Official source 1 ↗ · Official source 2 ↗
- Released training checkpoints. The flagship publishes intermediate CPT, cooldown, FTPO and SFT checkpoints as repository branches, plus training logs. Official source 1 ↗
Pricing & free limits
All sizes are downloadable from Hugging Face. Running costs depend on your hardware: the 375B flagship's reference deployment is 16 NVIDIA H100 GPUs (or one 8-GPU H200 node), while smaller models have GGUF and MLX builds for local use. Official source 1 ↗ · Official source 2 ↗
Billing details
Billing details were not established by the checked official sources.
Free access & limits
Free access and its limits were not established by the checked official sources.
Best for & limitations
Building Arabic and English assistants and agents that reason and call tools, for example answering an Arabic question that needs a live lookup via a function call. Official source 1 ↗
Running an Arabic-capable model on your own hardware, from a laptop-sized GGUF or MLX build of the smaller models up to multi-GPU serving of the flagship. Official source 1 ↗ · Official source 2 ↗
Research and experimentation on Arabic language models, using the published intermediate checkpoints and training logs. Official source 1 ↗
- Arabic and English only. Applications that need reliable performance in other languages are listed as out of scope. Official source 1 ↗
- Can be inaccurate or biased. IFM warns the model can produce made-up content and reflect training-data biases, and should not make medical, legal, financial or safety-critical decisions without human oversight. Official source 1 ↗
- Flagship needs datacenter GPUs. The 375B flagship's reference deployment uses 16 NVIDIA H100 80 GB GPUs across two nodes, or a single 8-GPU H200 node. Official source 1 ↗
- Technical report and code not yet out. The technical report and code repository are listed as coming soon. Official source 1 ↗
Submit a tool
