SSyncDev

Jais-3-3.7B

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Jais-3-3.7B is a small dense Arabic-English model in IFM's Jais 3 family with a 512K context, built for cheap single-GPU agents and strict instructions.

Key specifications

Specification score
53.4
Context window
524.29K
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Inputs and outputs

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

  • Text input

Produces

  • Text output

Capabilities

ReasoningFunction callingOpen weights

What is Jais-3-3.7B?

A 4B-class model scoring 41.2 on Terminal Bench 2.1 should make you look twice. That is what IFM reports for Jais-3-3.7B, the small dense member of Jais 3, and it is why this size is the interesting one for Arabic-language agents on a budget.

Released by MBZUAI's Institute of Foundation Models and G42's Inception42 under Apache-2.0, it runs on a single GPU and keeps the family's 512K-token context.

Head to head in its weight class

IFM compared it with four models of similar size. In every row of its table, the 3.7B came out on top:

Benchmark Jais-3-3.7B Qwen3.5-4B Gemma 4 E2B
Arabic MMLU 83.3 75.2 69.4
Arabic IFBench (strict) 92.6 53.7 37.6
Dialect ID (QADI) 68.8 37.4 44.4
Terminal Bench 2.1 41.2 25.8 0.4
BrowseComp-Plus 64.8 50.6 16.4
GDPVal 54.5 15.0 3.2

These are the publisher's numbers on its own harness. The size of the gaps on agentic tasks is unusual enough that you should reproduce one before you believe it.

One footnote on size. The name says 3.7B and the active count is 3.7B, but the repository metadata reports about 5B parameters stored. Budget memory for the larger figure.

A deployment sketch

This is the size you can afford to run per tenant or per region.

  • Hardware: one H100 at tensor-parallel 1 is the reference; the FP8 build (IFM/Jais-3-3.7B-FP8) cuts memory further.
  • Engines: vLLM 0.30.0+ or SGLang 0.5.20+, with the k2_horizon reasoning and tool-call parsers enabled.
  • Laptops: GGUF and 4-bit, 8-bit or bf16 MLX builds exist for local testing.
  • Requests: reasoning effort high on every call, and room for at least 32,768 output tokens. Thinking lands in reasoning_content, the answer in content.
  • Tools: xml by default, json or xml_typed on request.

Long outputs cost latency. If you are building a customer-facing Arabic assistant, measure the thinking overhead on short turns before you commit.

Where to point it

Arabic customer workflows that need a few tool calls: checking an order, filling a form, searching a knowledge base, replying in the customer's dialect. The dialect-identification score suggests it notices which Arabic it is reading, which matters for tone. It will not replace a frontier model for hard reasoning.

Timing is slightly fuzzy. The checkpoints appeared on 29 September 2026, while the family's public launch is reported as 15 October.

Testing it

Wire it into your actual tool schema and run a hundred recorded conversations end to end. Track task completion, tool-call validity and median turn latency on your GPU. Run the same set through Gemma 4 E2B and Qwen3.5 9B as references. Other options are under text generation.

How to evaluate Jais-3-3.7B

Compare the model on the work you need to complete.

Tasks and inputs

Jais-3-3.7B is listed for Text generation, Translation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

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.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

The reference setup is a single H100 at tensor-parallel 1; FP8, GGUF and MLX builds are also published.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Tencent

Hy4 Flash

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.

1.05M context

Alibaba

Qwen4 Max

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.

1M context

Sarvam

Sarvam-105B-1711

Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.

— context

IBM

Granite 4.1 8B LLMON Preview

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.

— context

NVIDIA

Nemotron-Labs Sigma 8B

Nemotron-Labs Sigma 8B is NVIDIA's open continuous diffusion language model, denoising 64-token blocks in parallel; English, GPU-only, 8K context.

8.19K context

Alibaba

Qwen4 Max Preview

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.

1M context