Model

Jamba2 Mini

OPEN

AI21 Labs · Jamba family · released Jan 8, 2026

Newest Jamba generation (successor to Jamba 1.5 Large/Mini already in corpus, dated 2024-08-22), built for enterprise RAG/grounding reliability. Released alongside a 3B dense sibling.

ReasoningCodingVisionFunction callingTool useAgentic

1933.4

Elo · unrated

Parameters

52B

Active params

12B (MoE)

Context

256K tokens

Architecture

SSM-Transformer hybrid (Mamba + Transformer) Mixture-of-Experts, 52B total / 12B active

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

FACTS GroundingKnowledge57.0%#12

best: Gemini 2.5 Flash-Lite (Thinking) · 86.8%

IFEvalReasoning97.0%#2

best: Gemma 4 26B A4B · 98.5%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Jamba family

Elo progression across releases

Jamba2 MiniJan 20261933.4

Resources

HFai21labs/AI21-Jamba2-Mini DOCShttps://www.ai21.com/blog/introducing-jamba2/APIAvailable via AI21 Labs API

API price weights · each benchmark row carries its own source badge (see methodology)