AI21 Labs (ai21.com) builds language models and the tooling around them, and now sells post-training work on customer data alongside its Jamba model family and the Maestro agent system. Its October 2026 homepage frames the company as "Lab-as-a-service" post-training rather than a place to rent a model by the token.
What AI21 Labs does today
The homepage pitch is owning rather than renting: "Renting got you started. Owning gets you ahead." The delivery mechanism is people, not just an API. AI21's Forward-Deployed Researchers (FDRs) post-train specialized models on a customer's own data, and the site carries customer quotes reporting the resulting gains — for example, Emet by Deel reports a +20.7 point post-training gain on payroll questions.
The products menu lists two entries: Post-Training and Intelligent Gateway. Alongside those, AI21 maintains two developer-facing lines: the Jamba family of open models and AI21 Maestro for building knowledge agents.
Jamba models use a hybrid Mamba-Transformer (SSM-Transformer) architecture. The documented context window is 256K tokens, which AI21 describes as the longest among open models. Weights are downloadable, input and output are text only, and nine languages are officially supported: English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic and Hebrew.
Current Jamba models
| Model | Total / active parameters | Context | Version / snapshot | API endpoint |
|---|---|---|---|---|
| Jamba Large | 398B total (94B active) | 256K | 1.7, snapshot 2025-07 | jamba-large, pointing to jamba-large-1.7-2025-07 |
| Jamba Mini | 52B total (12B active) | 256K | 2 (Jamba2 Mini), snapshot 2026-01 | jamba-mini, pointing to jamba-mini-2-2026-01 |
| Jamba2 3B | 3B | 256K | 2, snapshot 2026-01 | None — intended for on-device / local use |
The model strings developers search for are built from three parts. In jamba-mini-2-2026-01, jamba-mini is the model size, 2 is the major version, and 2026-01 is the snapshot date. jamba-large-1.7-2025-07 follows the same pattern for the Large line. The undated endpoints jamba-mini and jamba-large are aliases that currently point at those snapshots; the docs recommend calling dated versions because an undated alias moves when a new snapshot ships.
Older snapshots are retired on published dates: Jamba Mini 1.7 on 2026-02-01, Jamba Large and Jamba Mini 1.6 on 2025-08-03, and Jamba Large and Jamba Mini 1.5 on 2025-05-06. Jamba2 3B has no hosted API endpoint; it is meant for on-device and local use.
A separate small model, Jamba Reasoning 3B, was released on October 8, 2025 as a compact open-source reasoning model that runs locally through llama.cpp or LM Studio.
AI21 Maestro
Maestro launched on July 8, 2025 as a system for building knowledge agents. It covers retrieval-augmented generation, execution planning, and Validated Output, which enforces user-defined requirements such as format and tone.
The system has been extended since launch. On August 7, 2025, Mistral models became available inside Maestro. On September 18, 2025, Requirements were made to work with both File Search and Web Search. On November 27, 2025, Together AI open-source models were added.
Data Connectors arrived on September 16, 2025 for enterprise sources: Amazon S3, Box, Confluence, Dropbox, Google Drive, Browse Website, Intercom, OneDrive, Microsoft Teams, Notion, ServiceNow, SharePoint and Zendesk.
Deployment and security
Because Jamba weights are downloadable, AI21 supports private deployment in a VPC, on-premises, or in a hybrid setup. The Jamba 1.5 paper describes ExpertsInt8 quantization, which lets Jamba-1.5-Large run at 256K context on 8×80GB GPUs (that paper covers the now-deprecated 1.5 generation, but it shows the hardware scale involved in self-hosting the Large model).
On compliance, AI21 states SOC 2 and ISO 27001, ISO 27017 and ISO 27018.
Pricing
| Item | Price | Notes |
|---|---|---|
| Jamba Mini | $0.20 / 1M input tokens, $0.40 / 1M output tokens | Pay As You Go |
| Jamba Large | $2 / 1M input tokens, $8 / 1M output tokens | Pay As You Go |
| Free trial | $10 in credits, 7 days | No credit card required |
| Pay As You Go | Usage-based | Unlimited seats |
| Custom Plan | Quoted | Volume discounts, premium rate limits, private cloud hosting, priority support, dedicated account manager |
For estimating spend, AI21 says an average token is roughly 1 word or 6 characters of English text. The Custom Plan is where private cloud hosting sits, so self-hosting and volume pricing are negotiated rather than listed.
Who it is for
The post-training offer fits enterprises with proprietary data and a domain where a general model underperforms — payroll questions at Deel are the example AI21 publishes. The Jamba line fits teams that need a 256K context window, one of the nine supported languages, or weights they can run inside their own network. Maestro targets teams assembling knowledge agents over the connectors listed above. Jamba2 3B and Jamba Reasoning 3B suit on-device and local work.
Limitations to know
The model details state a knowledge cutoff of August 22, 2024. Jamba is text in, text out only — no image or audio input. AI21 notes the models perform best on English and can be inconsistent. Because undated endpoints move, the docs recommend pinning dated API versions in production. Jamba2 3B has no hosted API endpoint.
