Alpha
Contact
Use · Jais 3

Is it good enough?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input C ยท 57.6/100

This page is a projection of the one entry record, the Reliability factor that Use covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Capabilities & modalities

Jais 3 is a family of text reasoning and agentic models for Arabic and English. The report says they plan multi-step tasks, call tools and Model Context Protocol servers, operate a terminal, search the web, and read, edit and produce PDF, spreadsheet and presentation files in Arabic (ev-report-release). They also follow Arabic-specific instruction constraints such as hamza, diacritics and morphology.

Sizes: dense 0.9B, 3.7B, 7B and 32B, a 36B mixture of experts with Mixture-of-Value Attention (about 4B active, 100 experts in the config) and a 375B mixture of experts (about 23B active, 192 experts) (ev-hub).

The report says the 3.7B and 7B lead every agentic benchmark against models of their size, and that the 32B and 375B trail Qwen3.8 models on most of them (ev-report-eval). All figures are the publisher's own. Speech is a separate model (Jais-ASR) and is not covered.

Context window & long-context behaviour

The configuration sets max_position_embeddings to 524,288 for the 3.7B, 7B, 32B, MoVA-36B-A4B and 375B-A23B, and to 131,072 for the 0.9B (ev-hub). The report says the long context comes from K2 Horizon mid-training (ev-report-lineage).

The report scores the 3.7B at 68 on AA-LCR-Ar, an Arabic long-context benchmark with inputs of about 100K tokens, and says most Arabic-focused baselines could not run it (ev-report-eval). No result is reported near 512K tokens. Test your own long inputs and plan KV-cache memory (see implement.md).

Prompt format & chat template

Each repository ships chat_template.jinja. Reasoning depth is chosen per request through chat_template_kwargs, and the card says all reported results use reasoning_effort="high" (ev-card-32b). Low effort is not recommended except for research on reasoning effort.

The template accepts tool_call_format (json, xml or xml_typed, default xml) and tool_presentation_format (json, xml or markdown, default markdown). It raises an error for the misspelt names tool_presentation, tool_calling_format and tool_format. Enable the k2_horizon reasoning parser for chat and add the k2_horizon tool-call parser for agent use. Leave both off for plain completion.

Sampling from the card: temperature 0.2 by default, 1.0 for English thinking and 0.7 for Arabic thinking, with top_p 0.95 and top_k 20.

Language coverage

The cards tag Arabic and English. Arabic coverage is Modern Standard Arabic, the regional dialects and Arabizi, the Latin-script form of informal Arabic (ev-card-32b, ev-report-dialect).

Dialect evidence in training has four parts (ev-report-data, ev-report-dialect). Parallel dialect text comes from English conversations translated by GLM-5 and Jais 2 and judge-filtered. Arabizi covers Gulf, Egyptian, Levantine and Darija forms (25.1B tokens). Country-knowledge questions cover the 22 Arab League states, about 70% in the country's dialect. Translation and transliteration data covers seven dialects. The public sample's web subset is labelled 89.5% MSA, 6.7% Moroccan, 2.5% Egyptian and 1.0% Najdi (ev-data-sample).

Dialect evidence in evaluation: QADI dialect identification, Dialectal-Arabic-MMLU in Egyptian, Saudi, Moroccan, Syrian and Emirati, FLORES+ translation across eight dialects, and Darija and Egyptian transliteration. All runs are the publisher's own. Languages other than Arabic and English are out of scope on the cards.

Function / tool calling

Tool calling goes through the k2_horizon tool-call parser on vLLM or SGLang, with an XML default and JSON and typed-XML options (ev-card-32b). The card includes an Arabic function-calling example with a loop that executes tool calls.

The report evaluates agentic tasks on eleven benchmarks and says only two are in Arabic. Coding, tool use and deep search are tested in English only (ev-report-eval). Sandbox every tool and test the parser before an unattended run.

How this scores

The ownership factor this domain covers, drawn from the one entry record.

3

ReliabilityIs it reliable and good enough for the job?

Weak

Safety scores 2: every result is the publisher's own, there is no guard model, and no test covers agent actions. Operational scores 3 (ev-report-safety, ev-card-32b, ev-hub).

How this scores (AOI sub-dimensions)
Operational3/5how practical it is to run, serve and maintain in productionThe cards give validated vLLM, SGLang and Transformers recipes, parsers for reasoning and tool calls, tested GPU layouts and a 512K context for five of the six sizes (ev-card-32b, ev-hub).
Safety2/5whether misuse risks are evaluated and guardrails are providedThe models are safety tuned and the report publishes results on four benchmarks in two languages (ev-report-safety).
What this means for adoptionYou own the weights, every stage checkpoint, the training and evaluation code and a 23.8B-token data sample, all under Apache-2.0, so you can run and change the whole pipeline from any stage. You do not own a deployable safety case: safety results are the publisher's own, no guard model exists and agent actions are untested. Ownership moves to substantial if an independent evaluation lifts safety to 3 or more.

Sources

The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.

Model cardread2026-11-19
The 32B card declares license apache-2.0 in its front matter and has no License section.
Model cardread2026-11-19
The 0.9B, 3.7B, 7B, MoVA-36B-A4B and 375B-A23B cards repeat the 32B licence declaration, intended-use text and limitations.
Licenceread2026-11-19
Apache-2.0 is declared in the front matter of all six weights cards and in section 7.1 of the technical report.
Technical_reportread2026-11-19
The report says each Jais 3 model starts from the final post-trained K2 Horizon checkpoint of the same size, keeps its architecture and tokenizer, and is adapted by continual pretraining, SFT, FTPO and RLVR.
Technical_reportread2026-11-19
The report lists released weights, stage checkpoints, training and RL code, a training-data sample, run configurations, W&B logs, audited benchmarks and evaluation code (Table 41).
Technical_reportread2026-11-19
The report names each source with token counts: ArabicTxT360 web text (28.4B), literature collections (9.65B), translated reasoning (30.9B), parallel dialects (26.9B), Arabizi (25.1B) and 11.9B tokens of partner-supplied Arabic books with no named source.
Technical_reportread2026-11-19
The report says GLM-5, Jais 2, Qwen3.8-27B, Qwen3.6-27B, gemma-4-31B-it and MiniMax-M3 generated or judged parts of the training data.
Technical_reportread2026-11-19
Section 6.5 reports publisher-run results on AraSafe (1,254 harmful prompts, above 92% refusal for every size), Arabic Regional Safety, HarmBench and XSTest, scored with gemma-4-26B-A4B-it.
Technical_reportread2026-11-19
The report says Jais 3 leads or matches open models of its size on Arabic benchmarks.
Technical_reportread2026-11-19
Training covers Gulf, Egyptian, Levantine and Darija Arabizi, country knowledge for 22 states, and translation for seven dialects.
Documentationread2026-11-19
All six weights repositories are ungated, hold safetensors only with LFS sha256 values, and report a completed hub scan with no flagged files.
Documentationread2026-11-19
Each weights repository has branches CPT, Cooldown, FTPO and SFT besides main.
Documentationread2026-11-19
The IFM organisation hosts FP8, MLX and GGUF builds of the Jais-3 models, each tagged Apache-2.0.
Documentationread2026-11-19
IFM/Jais-3-Data is an ungated 23.8B-token sample, about 10% of CPT and SFT tokens, in four Arabic subsets licensed ODC-BY or CC BY 4.0.
Documentationread2026-11-19
The xllm, RL360 and gen-eval-arabic repositories are public with Apache-2.0 LICENSE files.
Documentationread2026-11-19
The JaisVoice README dates the Jais 3 release to 2026-10-15.
Model cardread2026-11-19
The K2-Horizon-32B card declares Apache-2.0, says "Fully open" and links code and data.
Technical_reportread2026-11-19
The K2-Horizon abstract describes six fully open models, about 10T synthetic pre-training tokens and a release that follows the LLM360 definition of open source (publisher claims).
Licenceread2026-11-19
The MiniMax Community License grants use for non-commercial purposes and sets conditions on commercial use of the software and its derivative works.
Model cardread2026-11-19
The four Jais-3 Uno adapter repositories carry no licence tag and no licence statement.
Documentationunverified2026-11-19
The ifm.ai blog and tech-report pages return HTTP 403 (Cloudflare challenge), so their text is unavailable.