Model
Model explorer

AI21 Jamba

OPEN
AI21 Labs · AI21 Jamba family · released Mar 28, 2024

First production-grade hybrid Mamba-Transformer MoE base model, 12B/52B active/total, 256K context fitting 140K tokens on one 80GB GPU.

ReasoningCodingVisionFunction callingTool useAgentic
302.4
Elo · rank #346
Parameters
52B
Active params
12B (MoE)
Context
256K tokens
Architecture
Hybrid Mamba-Transformer MoE (Jamba)
License
Apache-2.0
Languages
—
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
ARC-ChallengeReasoning64.4%#67
best: Llama 3.1 405B · 96.9%
ARC-EasyReasoning73.5%#39
best: Phi-3-medium (14B) · 97.7%
BIG-Bench HardReasoning45.4%#102
best: ERNIE 4.5 300B-A47B · 94.3%
GPQA DiamondReasoning2.5%#289
best: GPT-5.6 · 94.6%
GSM8KMath59.9%#115
best: Llama 3.1 405B · 96.8%
HellaSwagReasoning87.1%#19
best: Claude 3 Opus · 95.4%
HumanEvalCoding29.3%#157
best: Claude Opus 4.5 · 99.4%
IFEvalReasoning20.3%#155
best: Gemma 4 26B A4B · 98.5%
MATH-500Math1.6%#206
best: GPT-5 · 99.4%
MMLU-ProKnowledge16.6%#175
best: Claude Fable 5 · 91.5%
MMLUKnowledge67.4%#164
best: OpenAI o3 · 92.9%
PIQAReasoning83.2%#16
best: GPT-4o mini · 93.1%
TruthfulQAKnowledge46.4%#62
best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%
WinoGrandeReasoning82.5%#21
best: PaLM 2 · 90.9%
Run it locally
VRAM @ Q4
29.5 GB
VRAM @ FP16
104 GB
Fits on (Q4)
RTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
Throughput data unavailable.
Quantizations
GGUF Q4 · GPTQ
AI21 Jamba family
Elo progression across releases
AI21 JambaMar 2024302.4AI21 Jamba 1.5 LargeAug 2024777.9AI21 Jamba 1.5 MiniAug 2024553.6
Resources
HFai21labs/Jamba-v0.1DOCSdocs.ai21.com
API price weights · each benchmark row carries its own source badge (see methodology)