Model
Model explorer

Jais 70B

OPEN
Inception (G42) · Jais family · released Aug 5, 2024

Largest Jais model, adapted from Llama-2-70B; launched alongside 20+ other Jais variants (590M-70B) in the Aug 2024 family release.

ReasoningCodingVisionFunction callingTool useAgentic
112.8
Elo · rank #380
Parameters
70B
Active params
70B (dense)
Context
4K tokens
Architecture
Dense Llama-2-70B architecture adapted with Arabic-expanded tokenizer (32K new Arabic tokens), RoPE + GQA
License
Apache 2.0
Languages
2+
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
ARC-ChallengeReasoning54.9%#109
best: Llama 3.1 405B · 96.9%
HellaSwagReasoning82.7%#61
best: Claude 3 Opus · 95.4%
MMLUKnowledge38.7%#259
best: OpenAI o3 · 92.9%
OpenBookQAReasoning44.4%#39
best: Claude 1 · 90.8%
PIQAReasoning81.2%#34
best: GPT-4o mini · 93.1%
RACE-HReasoning42.9%#19
best: Claude 3 Opus · 92.9%
Social IQaReasoning52.9%#17
best: Apple DCLM-Baseline 7B · 82.9%
TruthfulQAKnowledge44.0%#67
best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%
WinoGrandeReasoning75.7%#61
best: PaLM 2 · 90.9%
Run it locally
VRAM @ Q4
40 GB
VRAM @ FP16
140 GB
Fits on (Q4)
M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
Throughput data unavailable.
Quantizations
GGUF Q4
Jais family
Elo progression across releases
Jais-13BAug 2023-409.7Jais 70BAug 2024112.8
Resources
HFinceptionai/jais-adapted-70b
API price weights · each benchmark row carries its own source badge (see methodology)