Tasks and inputs
Jais-3-32B 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-3-32B is the large dense model in IFM's open Jais 3 family: Arabic-English, 512K context, served on two H100s, strong on Arabic instruction following.
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
Two H100s, one model, no expert routing to tune. That is the operating appeal of Jais-3-32B, the large dense member of Jais 3 from MBZUAI's Institute of Foundation Models and G42's Inception42. It is a 32B-class decoder with a 512K-token context window, released under Apache-2.0.
Dense matters if your team has never run a mixture-of-experts model in production. There is less to go wrong.
The honest reading of IFM's table is a split decision. Against Gemma 4 31B, Qwen3.8-27B and three other 27B-34B models, the 32B:
Leads on: Arabic IFBench (88.7), Arabic IFEval (80.0), Arabic Legal (73.4), dialect identification (70.6), MCPMark (66.1), BrowseComp-Plus (71.9), and SimpleQA Verified in Arabic (28.1).
Trails on: Arabic MMLU (88.6 against Gemma 4's 89.5), Belebele, AraGen, and most agentic coding. Terminal Bench 2.1 is 63.7 where Qwen3.8-27B reaches 79.8; SWE-Bench Verified is 72.4 against 73.5.
So it is an instruction-following and Arabic-domain model first and a coding agent second. Publisher-reported throughout.
| Setting | Value |
|---|---|
| Hardware | One H100 node, tensor-parallel 2, FlashAttention-3 |
| vLLM | 0.30.0+, --trust-remote-code, --model-impl vllm, --dtype bfloat16 |
| SGLang | 0.5.20+, --tp 2, --attention-backend fa3 |
| Parsers | k2_horizon for reasoning and tool calls |
| Output budget | 32,768 tokens or more |
| Reasoning | reasoning_effort: "high" on every request |
| Builds | BF16, FP8 (IFM/Jais-3-32B-FP8), GGUF, MLX 4/8-bit and bf16 |
| Licence | Apache-2.0 |
The repository also exposes training-stage branches (CPT, Cooldown, FTPO, SFT), which you can select with --revision. Useful if you want to fine-tune from a less opinionated checkpoint.
One size discrepancy: metadata reports about 35B stored parameters against the 32B name. Plan memory for the larger figure.
Teams building Arabic legal, government or compliance assistants, where following a long, strict instruction set matters more than writing code. It is also the size where the dense Uno speed-up adapter is most valuable, since the model is large but still dense.
If cost per token is the bigger worry, compare the sparse Jais-3-MoVA-36B-A4B sibling, which activates only 4B parameters per token on the same two-GPU node.
If you track release dates, note that the weights were uploaded on 29 September 2026 and the launch is reported for 15 October.
Build two test sets: thirty long, rule-heavy Arabic instructions from your domain, and thirty agentic tasks with your real tools. Run both against this model and Gemma 4 31B. Score instruction compliance line by line on the first set and task success on the second. If the 32B wins the first and loses the second, you have your routing rule. More options live under text generation.
Compare the model on the work you need to complete.
Jais-3-32B 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.