Jamba Mini 2
Language Available Full comparison ↗ ai21/jamba-mini-2 · by AI21 Labs
· mixture-of-experts
Pricing — 1 offering(s)
Input tokens
- $0.20 / 1M tokens (input) Current 2026-01-01 → present
Output tokens
- $0.40 / 1M tokens (output) Current 2026-01-01 → present
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Jamba Mini 2 fits into a cost-aware routing setup
See how →Capability profile
How Jamba Mini 2 rates across core capability dimensions, with the task-level evidence behind each rating.
256,000 tokens — the same long-context window as Jamba Large, unusual for a model this small and cheap.
AI21 reports category-leading results for its size on IFBench, IFEval, Collie, and FACTS — instruction-following and grounded-factuality benchmarks aimed at enterprise tasks.
Structured JSON output and function calling supported (AI21 docs / HF model card).
Small model — not for hard multi-step reasoning; pair with a larger model or route reasoning-heavy steps elsewhere.
Same officially-supported language set as Jamba Large (English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic, Hebrew).
AI21 claims up to ~2.5x faster inference than comparable models; the fast tier of the Jamba line.
$0.20 / 1M input, $0.40 / 1M output via AI21's API — very cheap for a 256K-context model, and Apache-2.0 for self-hosting.
Operator guidance
The go-to for cheap, fast, long-context work that is mostly extraction, formatting, grounded summarisation, or instruction-following rather than open-ended reasoning. Escalate to Jamba Large 1.7 (same architecture, higher ceiling) or a flagship when answer quality on hard queries matters. Among cheap small models, its distinguishing features are the full 256K context and the Apache-2.0 license.
Use cases
- High-volume grounded summarisation, extraction, and classification over long documents where the task is constrained and cost matters
- Instruction-following / structured-output pipelines (JSON, form-filling, routing) at low cost
- Long-context RAG on a budget, or as the cheap tier in a two-model cascade behind Jamba Large or a flagship
- Self-hosted deployment under Apache 2.0 for data-control or offline requirements
Limitations
- Weak on hard reasoning and on coding — a small model, positioned for constrained enterprise tasks
- Hybrid SSM-Transformer fine-tuning/tooling ecosystem is smaller than for standard Transformers
- AI21's pricing page lists only a generic 'Jamba Mini' rate; the versioned mapping (jamba-mini -> jamba-mini-2-2026-01) is from AI21's docs (see registry entry); effective_from uses the alias month, not a confirmed exact day
- Capability ratings are qualitative from AI21's materials and the HF model card; the benchmark claims (IFBench etc.) are AI21-reported