AI21 Jamba¶

Research cut-off: 2026-07-22 Status: Phase 3, not yet researched. Every section below states precisely what must be sourced and under what conditions it may be recorded. No specification, price, benchmark result, or capability claim appears in this file, because none has been verified. Sections are never deleted: where the evidence does not exist, the section will read Not publicly disclosed or Insufficient independent evidence.

Required source set before this profile may be written, each recorded in data/sources.csv with an access date: the official model card or technical report; the API reference; the pricing page; the licence text or terms of service; the official code repository; and the official weights repository where the weights are public. Independent evaluations are recorded in data/benchmarks.csv and referenced here rather than restated.

Overview¶

Scope. One paragraph identifying the family, its position in the market, and what distinguishes it. Required source: the provider's own model documentation index.

Organization¶

Scope. The organisation that develops and serves the models, its corporate status, and its principal serving partners. Required source: official corporate and documentation pages.

Current model lineup¶

Scope. One row per model identifier with release date and status (generally available, preview, deprecated, retired). Generated from data/models.csv. Required source: the provider's model list and deprecation policy pages.

Architecture¶

Scope. Architecture family as disclosed, never as inferred from behaviour. Required source: technical report or model card.

Total parameters¶

Scope. Disclosed total parameter count, or 'Not publicly disclosed'. Estimation from benchmark behaviour or inference price is prohibited.

Active parameters¶

Scope. Disclosed parameters used per token. Equals the total for a dense model. 'Not publicly disclosed' otherwise.

Dense or mixture-of-experts¶

Scope. The disclosed layer structure, with expert count and routing where published.

Training and post-training¶

Scope. Pretraining scale, data description, and post-training method as disclosed. Record which elements are undisclosed rather than omitting the question.

Modalities¶

Scope. Accepted input modalities and produced output modalities, per model identifier. Required source: model card or API reference.

Context window¶

Scope. Advertised context length per model identifier, with a date, and any independently measured effective context, clearly separated.

Maximum output¶

Scope. Advertised maximum output length per model identifier, with a date.

Tokenizer¶

Scope. Tokenizer name or family where disclosed, and whether a token counting tool is published.

Reasoning modes¶

Scope. Whether extended reasoning is unavailable, caller-selectable, always applied, or offered as a distinct model; whether traces are visible; whether reasoning tokens are billed.

Tool use¶

Scope. Documented structured tool-calling support, including parallel calls and any constraints. Provider documentation, not measured performance.

Agent capabilities¶

Scope. Documented agentic features such as computer use or long-horizon loops. Measured agent performance belongs in the evidence sections below.

Coding performance¶

Scope. Recorded results on the coding benchmarks documented in this repository, each with harness, sampling policy, tool permissions, evidence grade, and date. Provider-reported results are labelled.

Scientific reasoning¶

Scope. Recorded results on knowledge and reasoning benchmarks, under the same conditions requirement.

Long-context performance¶

Scope. Recorded results on long-context benchmarks, with the measurement threshold stated for every effective-context figure.

Multimodal performance¶

Scope. Recorded results on multimodal benchmarks, per modality.

Factuality and hallucination¶

Scope. Recorded results on factuality benchmarks, with the grounding condition stated, since grounded and closed-book rates are different quantities.

Latency¶

Scope. Recorded latency measurements with percentile, concurrency, region, prompt and output length, and date. Vendor claims are labelled provider-reported.

Throughput¶

Scope. Recorded throughput measurements with concurrency and batching policy stated.

Token pricing¶

Scope. Input, cached input, and output prices in USD per million tokens with an effective date, cross-referenced to data/pricing.csv.

Caching¶

Scope. Cache mechanism, granularity, retention period, and discount, as documented.

Batch pricing¶

Scope. Batch discount and its conditions, as documented.

Hardware requirements¶

Scope. For open-weight models, the minimum accelerator configuration at a named precision. For API-only models, 'Not applicable'.

Quantization¶

Scope. Published quantization formats and any measured accuracy cost. Do not assert that a format is lossless without a measurement.

Memory footprint¶

Scope. Weight and KV cache memory at stated precision, sequence length, and batch size, computed with scripts/estimate_memory.py and labelled as an estimate.

Energy evidence¶

Scope. Published energy measurements stating all eleven required conditions, or 'Insufficient independent evidence'. No universal per-model figure is published.

Privacy and data retention¶

Scope. Data retention period, training-use default, opt-out mechanism, and regional processing, each as a dated snapshot with a URL.

Licensing¶

Scope. Exact licence name, its restrictions, and a link to the licence text. For API-only models, the terms of service governing use of outputs.

Strengths¶

Scope. Claims supported by Grade A or Grade B evidence recorded above, each naming the benchmark and conditions. No unqualified comparatives.

Limitations¶

Scope. Documented weaknesses, with the same evidence requirement as strengths.

Best use cases¶

Scope. Task classes for which the recorded evidence supports selection, with the evidence named. Not a recommendation absent a reader's evaluation.

Inappropriate use cases¶

Scope. Task classes for which the recorded evidence indicates the family is unsuitable, or for which licence or privacy terms preclude use.

Independent evidence¶

Scope. Grade A and Grade B sources for this family, listed separately from provider material so that a reader can see the independent record alone.

Provider-reported evidence¶

Scope. Grade C sources, listed separately and labelled. Included for completeness, never used alone to rank this family against another.

Open research questions¶

Scope. Questions this profile could not answer, and what would be required to answer them. Consolidated into chapter 21.

Sources¶

Scope. Numbered footnote definitions, each resolving to a source_id in data/sources.csv.