AI21 Labs (Jamba)
Listed at https://www.ai21.com
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | Jamba 1.5 Mini: $0.20/M tokens, Jamba 1.5 Large: $2/M tokens | per million tokens |
| Output | Jamba 1.5 Mini: $0.20/M tokens, Jamba 1.5 Large: $2/M tokens | per million tokens |
🤖 Supported Models (3)
✨ Pros
- ✓Jamba 1.5 series uses SSM-Transformer hybrid architecture with 398B total/94B active params and native 256K context
- ✓OpenAI-compatible API (/v1/chat/completions), zero SDK rewrite needed
- ✓Jamba 1.5 Mini priced at $0.20/M tokens, competitive with GPT-4o-mini for most workloads
- ✓Supports function calling and JSON mode
- ✓Distributed through AWS Bedrock, Azure AI Foundry, and NVIDIA NIM marketplaces
⚠️ Cons
- ×Limited model variety (Jamba series only, no Llama / Qwen / DeepSeek rehosting)
- ×China access requires stable proxy, no ICP-filed infrastructure
- ×Jamba 1.5 Large ($2/M) is more expensive than Llama 3.3 70B ($0.39-$0.88/M)
- ×Function calling reliability is weaker than OpenAI / Anthropic tier-1 vendors
- ×Third-party benchmark coverage (MMLU, HumanEval) is thinner than top closed-source models
🎯 Best For
Mid-to-large enterprises needing long context (256K) + high throughput; teams leveraging SSM's memory efficiency; multi-cloud distribution (Bedrock + Azure AI)
💰 Pricing & Plans
| Model | Input ($/M tokens) | Output ($/M tokens) | Best For |
|---|---|---|---|
| Jamba Mini | $0.20 | $0.40 | Lightweight general-purpose tasks, high-volume workloads |
| Jamba Large | $2.00 | $8.00 | Long-context (256K) reasoning, complex RAG, agentic workflows |
| Jamba Instruct | Same as underlying base | Same as underlying base | Instruction-tuned variants of Mini / Large for chat and assistants |
| AI21 Studio (Free Tier) | $0 (trial credits) | $0 (trial credits) | Sandbox access at studio.ai21.com for prototyping |
| Enterprise / On-Demand | Custom contract | Custom contract | Private deployments, custom SLAs, dedicated capacity |
🔧 API & Developer Experience
- •API Style: OpenAI-compatible /v1/chat/completions endpoint plus AI21's native /v1/chat and /v1/completion routes — existing OpenAI SDKs work by swapping the base URL to https://api.ai21.com/studio/v1.
- •Base URLs: Studio (managed): https://api.ai21.com/studio/v1 for direct API; AWS Bedrock, Azure AI Foundry, and NVIDIA NIM each expose Jamba under their own marketplace base URLs with the same schema.
- •SDK Availability: Official Python SDK (ai21) and TypeScript/Node client; community wrappers for Go, Java, Ruby. AI21's docs include curl examples and OpenAI SDK migration snippets.
- •Function Calling & JSON Mode: Tool-use (function calling) is supported on Jamba Mini and Jamba Large with structured JSON output — reliability trails OpenAI/Anthropic for complex multi-tool chains but covers single-tool schemas well.
- •Long Context (256K): Native 256K-token context window on both Jamba Mini and Jamba Large — no chunking required for long-document summarization, large codebase ingestion, or extended agentic memory.
- •Streaming & Sampling Controls: Server-Sent Events (SSE) streaming by default; full OpenAI-style sampling parameters (temperature, top_p, max_tokens, stop, presence/frequency penalty) on every model.
- •Distribution Channels: Available via AI21 Studio (direct), AWS Bedrock (fully-managed), Azure AI Foundry (enterprise), and NVIDIA NIM (self-hosted container) — same model weights, four procurement paths.
🧠 SSM-Transformer Hybrid Architecture (Jamba)
Jamba is the only major production LLM that fuses Mamba (a Selective State-Space Model) with Transformer attention in a single weights file. The hybrid blocks — alternating Mamba layers, a small number of attention layers, and a Mixture-of-Experts (MoE) feed-forward — give Jamba 1.5 Large 398B total parameters but only 94B active per token. This combination is what makes 256K native context affordable: Mamba's linear-time sequence handling keeps KV-cache and inference cost low, while the interleaved attention layers preserve the in-context recall that pure SSMs lose on long prompts. Jamba Mini and Large both ship under an Apache 2.0 license and are available as raw weights on Hugging Face, which has made them a popular backbone for self-hosted long-context pipelines and on-prem enterprise deployments where data residency rules out API-only providers.
🌐 China Access & Latency
AI21 Studio (api.ai21.com) is hosted in the US/EU and is not directly reachable from mainland China; the Studio signup flow also blocks Chinese payment methods on some plans. Chinese developers typically access Jamba via proxy or through one of three indirect routes: AWS Bedrock (ap-northeast-1 Tokyo or ap-east-1 Hong Kong give 100–250 ms latency), Azure AI Foundry (if the tenant has China-region peering), or NVIDIA NIM self-hosted inside a Hong Kong / Singapore / Tokyo colo. The Hugging Face raw-weights path is the cleanest workaround for fully China-resident deployments — pull the Jamba Mini or Large checkpoints and serve them on domestic GPUs (A100 / H800 / 国产替代) behind a vLLM / TensorRT-LLM endpoint with no cross-border traffic at all. Direct Studio latency from a stable proxy is typically 250–500 ms round-trip; raw-weights self-hosting under the same hardware delivers sub-50 ms first-token latency for Chinese end users.