Jais 3 32B Benchmarks, Specifications & Availability

Jais 3 32B

Explore Jais 3 32B from IFM / Inception42: published specifications, source-linked vendor benchmarks, independent evaluator coverage and recorded pricing when available.

Published specifications

Provider: IFM / Inception42
Access: Open weights
License: Apache-2.0
Context window: 512K
Total parameters: 32B
Active parameters: Not published
Released: 2026-10-15
Modalities: text
Family: Jais 3

Hugging Face · Announcement · Website

Model notes

Arabic/English text assistant and agent model by IFM at MBZUAI and Inception42, in collaboration with Cerebras. Final released main checkpoint adapted from the matching post-trained K2 Horizon; CPT/Cooldown/SFT/FTPO branches are training-stage alternatives, not these evaluation results. Dense model; all parameters active (not represented as MoE). Vendor-published size retained rather than a different raw tensor count. Native 512K context; press release calls it 512,000 tokens. All card results explicitly use high reasoning effort. Arabic tasks use vendor-audited cleaned question sets and generation-based scoring; preserve source-specific definitions. GDPval uses a 75% rubric threshold, WildClawBench English-only Mean@3, Apex-Agents a 420-task text-only loop-agent subset, SWE-bench offline; these are not unrestricted full-suite results. Technical report/code marked coming soon on the model card; source blog provides methodology and summary tables. No exact public OpenRouter entry found on 17 October 2026; pricing remains missing. Downloadable weights are open under Apache 2.0.

Compare Jais 3 32B with other models →Compare subscription plans Explore data coverage

Pricing · OpenRouter

No OpenRouter pricing is recorded for this model. Missing rates are not free.

Official / vendor benchmarks

Default headline records. Own-vendor, peer-vendor and third-party provenance remain visible in evidence; configurations may differ.

Official / vendor headline scores; expand evidence for every record
Benchmark / evaluator	Headline score	Evidence
AraGen v3 (3C3H, September 2025) (IFM Jais 3)	71%high effort	All 1 recorded result & sources 71% · raw 71 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: AraGen v3 (3C3H)09-2025. https://huggingface.co/inception42/Jais-3-32B
AraGen v3 — Correctness (IFM Jais 3)	81%high effort	All 1 recorded result & sources 81% · raw 81 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Correctness. https://huggingface.co/inception42/Jais-3-32B
AraGen v3 — Completeness (IFM Jais 3)	80.3%high effort	All 1 recorded result & sources 80.3% · raw 80.3 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Completeness. https://huggingface.co/inception42/Jais-3-32B
AraGen v3 — Conciseness (IFM Jais 3)	28.1%high effort	All 1 recorded result & sources 28.1% · raw 28.1 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Conciseness. https://huggingface.co/inception42/Jais-3-32B
AraGen v3 — Helpfulness (IFM Jais 3)	76.4%high effort	All 1 recorded result & sources 76.4% · raw 76.4 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Helpfulness. https://huggingface.co/inception42/Jais-3-32B
AraGen v3 — Honesty (IFM Jais 3)	79.9%high effort	All 1 recorded result & sources 79.9% · raw 79.9 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Honesty. https://huggingface.co/inception42/Jais-3-32B
AraGen v3 — Harmlessness (IFM Jais 3)	80.3%high effort	All 1 recorded result & sources 80.3% · raw 80.3 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Harmlessness. https://huggingface.co/inception42/Jais-3-32B
Arabic MMLU (IFM Jais 3; audited)	88.6%high effort	All 1 recorded result & sources 88.6% · raw 88.6 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arabic MMLU*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
Belebele (IFM Jais 3)	95%high effort	All 1 recorded result & sources 95% · raw 95 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: BelebeleAccuracy. https://huggingface.co/inception42/Jais-3-32B
AraSTEM (IFM Jais 3; audited)	95.2%high effort	All 1 recorded result & sources 95.2% · raw 95.2 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: AraSTEM*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
Arabic GSM8K (IFM Jais 3; audited)	94.3%high effort	All 1 recorded result & sources 94.3% · raw 94.3 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arabic GSM8K*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
GATmath (IFM Jais 3; audited)	98.5%high effort	All 1 recorded result & sources 98.5% · raw 98.5 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: GATmath*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
MedAraBench (IFM Jais 3; audited)	87%high effort	All 1 recorded result & sources 87% · raw 87 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: MedAraBench*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
Arabic Legal (HELM) (IFM Jais 3; audited)	73.4%high effort	All 1 recorded result & sources 73.4% · raw 73.4 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arabic Legal (HELM)*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
IslamicQA (IFM Jais 3)	88.7%high effort	All 1 recorded result & sources 88.7% · raw 88.7 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: IslamicQAAccuracy (mean of the 4 below). https://huggingface.co/inception42/Jais-3-32B
PaLM-X 2025-2 (IFM Jais 3; audited)	94.6%high effort	All 1 recorded result & sources 94.6% · raw 94.6 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: PaLM-X 2025-2*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
QIAS-1 (IFM Jais 3; audited)	95.9%high effort	All 1 recorded result & sources 95.9% · raw 95.9 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: QIAS-1*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
QIAS-2 (IFM Jais 3; audited)	85.5%high effort	All 1 recorded result & sources 85.5% · raw 85.5 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: QIAS-2*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
IslamicEval2025-1B (IFM Jais 3; audited)	78.8%high effort	All 1 recorded result & sources 78.8% · raw 78.8 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: IslamicEval2025-1B*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
Arab Culture (IFM Jais 3)	88%high effort	All 1 recorded result & sources 88% · raw 88 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arab CultureAccuracy (mean of the 3 below). https://huggingface.co/inception42/Jais-3-32B
ArabCulture (IFM Jais 3; audited)	88.2%high effort	All 1 recorded result & sources 88.2% · raw 88.2 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: ArabCulture*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
AraDiCE-Culture (IFM Jais 3; audited)	88.3%high effort	All 1 recorded result & sources 88.3% · raw 88.3 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: AraDiCE-Culture*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
Arabic Cuisine (IFM Jais 3)	87.4%high effort	All 1 recorded result & sources 87.4% · raw 87.4 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arabic Cuisine^Accuracy. Vendor says benchmark to be released; no public test-set availability assumed. https://huggingface.co/inception42/Jais-3-32B
Dialectal Arabic MMLU (IFM Jais 3; audited)	88.5%high effort	All 1 recorded result & sources 88.5% · raw 88.5 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Dialectal Arabic MMLU*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
Dialect Identification (QADI) (IFM Jais 3)	70.6%high effort	All 1 recorded result & sources 70.6% · raw 70.6 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Dialect Identification (QADI)Accuracy. https://huggingface.co/inception42/Jais-3-32B
Alyah (IFM Jais 3; audited)	83%high effort	All 1 recorded result & sources 83% · raw 83 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Alyah*Accuracy. Audited subset: items with invalid options, ambiguous questions or indefensible gold labels removed. https://huggingface.co/inception42/Jais-3-32B
Arabic MMLU with diacritics (IFM Jais 3)	87.9%high effort	All 1 recorded result & sources 87.9% · raw 87.9 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arabic MMLU with diacriticsAccuracy. https://huggingface.co/inception42/Jais-3-32B
FannOrFlop-MCQ (IFM Jais 3)	71.8%high effort	All 1 recorded result & sources 71.8% · raw 71.8 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: FannOrFlop-MCQAccuracy. https://huggingface.co/inception42/Jais-3-32B
Arabic Linguistics (IFM Jais 3)	94.3%high effort	All 1 recorded result & sources 94.3% · raw 94.3 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arabic Linguistics^Accuracy (mean of the 2 below). Vendor says benchmark to be released; no public test-set availability assumed. https://huggingface.co/inception42/Jais-3-32B
Diacritization (IFM Jais 3)	91.9%high effort	All 1 recorded result & sources 91.9% · raw 91.9 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: DiacritizationAccuracy. https://huggingface.co/inception42/Jais-3-32B
Parsing (IFM Jais 3)	96.7%high effort	All 1 recorded result & sources 96.7% · raw 96.7 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: ParsingAccuracy. https://huggingface.co/inception42/Jais-3-32B
FLORES200+LLM-as-a-judge (IFM Jais 3)	91.5%high effort	All 2 recorded results & sources 91.5% · raw 91.5 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: FLORES200+LLM-as-a-judge. https://huggingface.co/inception42/Jais-3-32B 91.5% · raw 91.5 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Arabic IFBench (IFM Jais 3)	88.7%high effort	All 1 recorded result & sources 88.7% · raw 88.7 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arabic IFBench^Accuracy (prompt-level strict). Vendor says benchmark to be released; no public test-set availability assumed. https://huggingface.co/inception42/Jais-3-32B
Arabic IFEval (IFM Jais 3)	80%high effort	All 1 recorded result & sources 80% · raw 80 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arabic IFEvalAccuracy. https://huggingface.co/inception42/Jais-3-32B
AA-LCR (Ar) (IFM Jais 3)	75%high effort	All 2 recorded results & sources 75% · raw 75 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: AA-LCR (Ar)Accuracy. https://huggingface.co/inception42/Jais-3-32B 75% · raw 75 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
HalluScore (IFM Jais 3)	67.3%high effort	All 1 recorded result & sources 67.3% · raw 67.3 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: HalluScoreAccuracy. https://huggingface.co/inception42/Jais-3-32B
SimpleQA Verified (Arabic) — weighted aggregate (IFM Jais 3)	28.1%high effort	All 1 recorded result & sources 28.1% · raw 28.1 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: SimpleQA Verified (Ar)Aggregate (2/3 F-score + 1/3 non-hallucination rate). Aggregate = two-thirds F-score + one-third non-hallucination rate; not ordinary accuracy. https://huggingface.co/inception42/Jais-3-32B
SimpleQA Verified (Arabic) — F-score (IFM Jais 3)	7.1%high effort	All 1 recorded result & sources 7.1% · raw 7.1 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: F-score. https://huggingface.co/inception42/Jais-3-32B
SimpleQA Verified (Arabic) — Non-hallucination rate (IFM Jais 3)	70.1%high effort	All 1 recorded result & sources 70.1% · raw 70.1 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Non-hallucination rate. https://huggingface.co/inception42/Jais-3-32B
Terminal Bench 2.1 (IFM Jais 3)	63.7%high effort	All 2 recorded results & sources 63.7% · raw 63.7 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Terminal Bench 2.1Accuracy (Avg@3). Accuracy Avg@3. https://huggingface.co/inception42/Jais-3-32B 63.7% · raw 63.7 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
GDPVal (IFM Jais 3)	75.5%high effort	All 2 recorded results & sources 75.5% · raw 75.5 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: GDPValAccuracy (@ 75% threshold). Task accuracy at a 75% rubric threshold, not Elo or mean rubric completion. https://huggingface.co/inception42/Jais-3-32B 75.5% · raw 75.5 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
MCPMark (IFM Jais 3)	66.1%high effort	All 2 recorded results & sources 66.1% · raw 66.1 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: MCPMarkAccuracy (Pass@1). https://huggingface.co/inception42/Jais-3-32B 66.1% · raw 66.1 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
WildClawBench (English text-only subset) (IFM Jais 3)	42%high effort	All 2 recorded results & sources 42% · raw 42 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: WildClawBench (English text-only subset)Mean Score@3. English text-only subset, Mean Score@3; not full multimodal suite or pass@3. https://huggingface.co/inception42/Jais-3-32B 42% · raw 42 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Apex-Agents (420 text-only subset, loop agent) (IFM Jais 3)	25.2%high effort	All 2 recorded results & sources 25.2% · raw 25.2 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Apex-Agents (420 text-only subset, loop agent)Pass@1 (Accuracy). 420-task text-only subset, loop agent. https://huggingface.co/inception42/Jais-3-32B 25.2% · raw 25.2 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
GAIA v2 lit (Ar) (IFM Jais 3)	51.5%high effort	All 2 recorded results & sources 51.5% · raw 51.5 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: GAIA v2 lit (Ar)Accuracy (Pass@1). https://huggingface.co/inception42/Jais-3-32B 51.5% · raw 51.5 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Arabic File Processing (IFM Jais 3)	71.8%high effort	All 2 recorded results & sources 71.8% · raw 71.8 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: Arabic File ProcessingAccuracy (@ 75% threshold). Task accuracy at a 75% rubric threshold, not Elo or mean rubric completion. https://huggingface.co/inception42/Jais-3-32B 71.8% · raw 71.8 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
SWE-Bench Verified (Offline) (IFM Jais 3)	72.4%high effort	All 2 recorded results & sources 72.4% · raw 72.4 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: SWE-Bench Verified (Offline)Accuracy (Pass@1). Offline protocol. https://huggingface.co/inception42/Jais-3-32B 72.4% · raw 72.4 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
BrowseComp-Plus (IFM Jais 3)	71.9%high effort	All 2 recorded results & sources 71.9% · raw 71.9 % Headline · high effort · Own vendor Source/record date: 2026-10-17 Harness: IFM/Inception42 final main checkpoint, high effort. IFM/Inception42 final-checkpoint evaluation. High effort. Arabic suites use generation-based evaluation; starred source tests use audited cleaned question sets, not the unfiltered public suite. Exact source label: BrowseComp-PlusAccuracy. https://huggingface.co/inception42/Jais-3-32B 71.9% · raw 71.9 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Jais 3 — Arabic average (IFM summary)	83.1%high effort	All 1 recorded result & sources 83.1% · raw 83.1 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Jais 3 — Knowledge & reasoning (IFM summary)	90.3%high effort	All 1 recorded result & sources 90.3% · raw 90.3 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Jais 3 — General knowledge & comprehension (IFM summary)	91.8%high effort	All 2 recorded results & sources 91.8% · raw 91.8 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/ 91.8% · raw 91.8 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 Figure-only summary printed on official model card; differs in some rounded values from launch blog. Blog summary remains headline; no averaging or reconstruction. https://huggingface.co/inception42/Jais-3-32B/resolve/main/assets/jais-3-32b-benchmarks.png
Jais 3 — STEM & math (IFM summary)	96%high effort	All 2 recorded results & sources 96% · raw 96 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/ 96% · raw 96 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 Figure-only summary printed on official model card; differs in some rounded values from launch blog. Blog summary remains headline; no averaging or reconstruction. https://huggingface.co/inception42/Jais-3-32B/resolve/main/assets/jais-3-32b-benchmarks.png
Jais 3 — Professional domains (IFM summary)	80.2%high effort	All 1 recorded result & sources 80.2% · raw 80.2 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Jais 3 — Arabic language & culture (IFM summary)	85.5%high effort	All 1 recorded result & sources 85.5% · raw 85.5 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Jais 3 — Islamic knowledge & Arab culture (IFM summary)	88.1%high effort	All 2 recorded results & sources 88.1% · raw 88.1 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/ 88.2% · raw 88.2 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 Figure-only summary printed on official model card; differs in some rounded values from launch blog. Blog summary remains headline; no averaging or reconstruction. https://huggingface.co/inception42/Jais-3-32B/resolve/main/assets/jais-3-32b-benchmarks.png
Jais 3 — Dialects (IFM summary)	80.7%high effort	All 2 recorded results & sources 80.7% · raw 80.7 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/ 80.7% · raw 80.7 % Alternative · high effort · Own vendor Source/record date: 2026-10-17 Figure-only summary printed on official model card; differs in some rounded values from launch blog. Blog summary remains headline; no averaging or reconstruction. https://huggingface.co/inception42/Jais-3-32B/resolve/main/assets/jais-3-32b-benchmarks.png
Jais 3 — Linguistics & literature (IFM summary)	84.7%high effort	All 1 recorded result & sources 84.7% · raw 84.7 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Jais 3 — Assistant capabilities (IFM summary)	67.8%high effort	All 1 recorded result & sources 67.8% · raw 67.8 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Jais 3 — Instruction following (IFM summary)	84.3%high effort	All 1 recorded result & sources 84.3% · raw 84.3 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Jais 3 — Factuality (IFM summary)	47.7%high effort	All 1 recorded result & sources 47.7% · raw 47.7 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/
Jais 3 — Agentic average (IFM summary)	60%high effort	All 1 recorded result & sources 60% · raw 60 % Headline · high effort · Own vendor Source/record date: 2026-10-17 As reported by IFM/Inception42. Final-checkpoint launch summary; Arabic overall/domain averages weight benchmark components, not displayed subdomain averages. Agentic average covers nine suites, BrowseComp replaces Plus at flagship scale; those averages are not identical test pools. Card definitions preserve audited sets/thresholds/subsets. https://ifm.ai/blog/jais3/

Independent evaluators

Evaluator harnesses are distinct from vendor measurements. Missing coverage is not a failed test.

No independent evaluator results are recorded for this model.

Read how we select and source scores or the comparison guide.