Model

AI21 Jamba 1.5 Large

OPEN

AI21 Labs · AI21 Jamba family · released Aug 22, 2024

Flagship open hybrid SSM-Transformer MoE, 94B/398B active/total, RULER-verified 256K context, function calling and JSON output.

ReasoningCodingVisionFunction callingTool useAgentic

777.9

Elo · rank #255

Parameters

398B

Active params

94B (MoE)

Context

256K tokens

Architecture

Hybrid Mamba-Transformer MoE (Jamba)

License

Jamba Open Model License

Languages

API price (in/out)

$2 / $8

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

ARC-ChallengeReasoning93.0%#16

best: Llama 3.1 405B · 96.9%

Arena-HardHuman preference65.4%#36

best: Qwen3-235B-A22B (Non-Thinking) · 96.1%

BIG-Bench HardReasoning65.5%#71

best: ERNIE 4.5 300B-A47B · 94.3%

BFCLAgents85.5%#6

best: SmolLM3 3B (No Thinking) · 92.3%

GPQA DiamondReasoning36.9%#233

best: GPT-5.6 · 94.6%

GSM8KMath87.0%#56

best: Llama 3.1 405B · 96.8%

HumanEvalCoding71.3%#81

best: Claude Opus 4.5 · 99.4%

IFEvalReasoning81.5%#88

best: Gemma 4 26B A4B · 98.5%

MMLU-ProKnowledge48.3%#135

best: Claude Fable 5 · 91.5%

MMLUKnowledge80.0%#91

best: OpenAI o3 · 92.9%

MMMLUKnowledge73.9%#37

best: Gemini 3.1 Pro · 92.6%

TruthfulQAKnowledge58.3%#28

best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%

WildBenchHuman preference48.5%#5

best: Mistral Small 3.2 (24B) · 65.3%

Run it locally

VRAM @ Q4

240.4 GB

VRAM @ FP16

796 GB

Fits on (Q4)

M3 Ultra 512GB

Too large to run on a single RTX 4090 even at Q4; requires multi-GPU or CPU offload.

Quantizations

GGUF Q4

AI21 Jamba family

Elo progression across releases

AI21 JambaMar 2024302.4 AI21 Jamba 1.5 LargeAug 2024777.9 AI21 Jamba 1.5 MiniAug 2024553.6

Resources

HFai21labs/AI21-Jamba-Large-1.5 DOCSdocs.ai21.com APIAvailable via AI21 Labs API

API price $2/$8 · each benchmark row carries its own source badge (see methodology)