Tasks and inputs
Jais-3-0.9B 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-3-0.9B is the smallest open Jais 3 model from MBZUAI's IFM and Inception42: a dense Arabic-English model with GGUF and MLX builds for local use.
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
How small can an Arabic assistant get before it stops being useful? Jais-3-0.9B is the Jais 3 family's answer: a dense model under a billion parameters, released by MBZUAI's Institute of Foundation Models with G42's Inception42 under Apache-2.0. It is built on IFM's K2 Horizon foundation models, like the rest of the family.
Its natural home is a laptop or a phone, not a GPU cluster.
Family-level material describes Jais 3 as a 512K-context line. This checkpoint is different. IFM describes this checkpoint as a 128K model, and the serving example sets a maximum length of 131,072 tokens via a YaRN scaling factor of 16 over an 8,192 base. Plan for 128K. If you need longer windows locally, a bigger sibling is the safer bet.
| Build | Use |
|---|---|
inception42/Jais-3-0.9B |
BF16 weights (Hub metadata rounds it to 1B) |
IFM/Jais-3-0.9B-GGUF |
llama.cpp and Ollama |
IFM/Jais-3-0.9B-MLX-4bit / -8bit / -bf16 |
Apple silicon |
Unlike the larger sizes, no FP8 build is listed for this one. That makes sense: FP8 targets datacentre GPUs, and this model is aimed elsewhere. On a server, a single H100 at tensor-parallel 1 is the reference setup.
The scores tell a clear story. Instruction following is strong for the size: 91.1 on Arabic IFBench (prompt-level strict) and 71.3 on Arabic IFEval. Arabic knowledge is respectable, with 70.4 on Arabic MMLU and 81.6 on Belebele. Factuality is the weak spot. HalluScore is 28.0 and SimpleQA Verified (Arabic) 27.1. All publisher-reported.
Translate that into product terms. Use it to follow instructions on text you give it: rewriting, tagging, classifying, extracting, short drafting in Arabic or English. Do not ask it to recall facts on its own. Pair it with retrieval, or route factual questions to a bigger model.
Every Jais 3 checkpoint is tuned to run with reasoning on. Pass reasoning_effort: "high" through chat_template_kwargs; low effort is described as for research only. Allow at least 32,768 output tokens, which is a lot for a phone, so cap thinking with care in on-device builds. Default sampling is temperature 0.2, top_p 0.95, top_k 20, rising to 1.0 (English) or 0.7 (Arabic) in thinking mode. Tool calls default to an XML format, with JSON also available, parsed by the k2_horizon parser in vLLM.
The weights are Apache-2.0. A dating note: the repositories went up on 29 September 2026, about two weeks before the widely reported 15 October launch, which we could not confirm on IFM's own press page.
Load the 4-bit MLX or GGUF build on the oldest device you support. Measure time to first token and tokens per second with thinking on, then with a tight token budget. Run fifty Arabic instructions from your real product and read the outputs for dialect handling. Comparable small multilingual options include Tiny Aya Global and Gemma 4 E2B.
Compare the model on the work you need to complete.
Jais-3-0.9B 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.