Tasks and inputs
Jais-3-7B 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-3-7B is the mid-size dense Arabic-English model in IFM's open Jais 3 family: 512K context, single-GPU serving and an optional Uno speed-up adapter.
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
IFM's post-training lifted the Arabic average of this model's K2 Horizon base from 56.4 to 79.6. That single number explains Jais-3-7B better than its size does. It is a general-purpose base turned into an Arabic-first assistant, released by MBZUAI's Institute of Foundation Models and G42's Inception42 under Apache-2.0.
It is dense, it has a 512K-token context, and it fits on one GPU.
Seven billion parameters puts it against models that are often larger. On IFM's own table:
| Benchmark | Jais-3-7B | Qwen3.5-9B | Gemma 4 12B |
|---|---|---|---|
| Arabic MMLU | 85.2 | 80.9 | 84.1 |
| Belebele | 92.9 | 89.2 | 93.6 |
| Arabic IFBench (strict) | 91.6 | 58.3 | 75.3 |
| FLORES200+ (judge) | 90.6 | 74.3 | 89.5 |
| AA-LCR (Arabic) | 65.0 | 68.0 | 49.0 |
| SWE-Bench Verified (offline) | 70.2 | 50.8 | 28.2 |
It does not win everything: the 12B Gemma 4 leads on Belebele and AraGen, and Qwen3.5 9B is ahead on Arabic long-context reasoning. The SWE-Bench figure is the one to verify yourself; 70.2 at this size is a big claim. All of these are publisher-reported.
Repository metadata lists about 9B stored parameters against the 7B name. Size your memory for the larger figure.
Dense Jais 3 sizes, this one included, have an optional "Uno" LoRA diffusion adapter. It drafts several tokens per pass, and the main autoregressive model verifies them, which IFM puts at about 1.9x faster on average. The catch is infrastructure: Uno needs a dedicated engine rather than stock vLLM. Worth it if you serve high volume; probably not for a prototype.
k2_horizon reasoning parser for chat, and add the tool-call parser for agents.reasoning_effort: "high" and allow 32,768 or more output tokens.xml tool calls unless your framework insists on JSON.Pick the 7B when Arabic quality has to be close to the big models but your budget is one GPU per replica. Pick the smaller 3.7B for many lightweight agents, and the 32B when one node with two GPUs is acceptable and you need the extra headroom on legal or instruction-heavy work.
One more caveat: the 15 October 2026 launch date is reported rather than confirmed first-hand, and the weights were public from 29 September. Context length, licence and tool support for this size follow the family-wide specification.
Translate twenty of your English prompts into the Arabic dialects you serve, run them with and without thinking, and grade fluency with native speakers. Then run a short agentic task set against Gemma 4 E4B as a cheaper baseline. More open options sit under text generation.
Compare the model on the work you need to complete.
Jais-3-7B 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.