Skip to main content

AI21Language modelOpen source

Jamba2-3B

Jamba2-3B has no overall index value yet because there are too few independent benchmark results.

PopularityRank 213 of 342 language models

34out of 100new calculation

usage data only, range 20 to 48

Summary

Jamba2-3B combines state-space and Transformer layers in a dense 3B model. AI21 targets answers grounded in supplied documents, including manuals and company policies. The compact version works without a reasoning model’s additional thinking process.

AI21 specifies 256K context tokens and releases the weights under Apache 2.0. Access is available through AI21 Studio and Hugging Face, with Jamba2-Mini joining the release as a larger MoE variant.

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 10 published benchmark results

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

MMLU Turkish, current 900-question test subset, zero-shot (Show measurement, test conditions, and source)
Source value
22.8
Metric
accuracy
Unit
%
Benchmark version
AYueksel/TurkishMMLU, current 900-question test subset
Category
multilingual reasoning
Direction
Higher is better
Sample
900
Source type
community-reported
Evaluator
serda-dev, Jamba2-3B-Turkish README benchmark comparison
Status
active
Retrieved at
2026-10-20
Methodology
Community-reported result. The README states these accuracy results were obtained 2026-10-14 under the same protocol. The measured row is “Orijinal Jamba2 (3B)”. The upstream immutable revision for that baseline is not stated. Dataset: AYueksel/TurkishMMLU, current 900-question test subset; zero-shot option-letter logits. This is not the full or five-shot MMLU leaderboard.
Context
The exact table row is “Orijinal Jamba2 (3B)”, distinct from the Turkish SFT and CPT rows. The source does not identify an immutable upstream revision for the original Jamba2 baseline. Do not transfer these scores to the Turkish fine-tunes.
Source serda-dev, Jamba2-3B-Turkish README benchmark comparison
Belebele Turkish tur_Latn/test accuracy (Show measurement, test conditions, and source)
Source value
49
Metric
accuracy
Unit
%
Benchmark version
facebook/belebele, tur_Latn/test
Category
multilingual reasoning
Direction
Higher is better
Sample
900
Source type
community-reported
Evaluator
serda-dev, Jamba2-3B-Turkish README benchmark comparison
Status
active
Retrieved at
2026-10-20
Methodology
Community-reported result. The README states these accuracy results were obtained 2026-10-14 under the same protocol. The measured row is “Orijinal Jamba2 (3B)”. The upstream immutable revision for that baseline is not stated. Dataset split: facebook/belebele, tur_Latn/test; all 900 questions.
Context
The exact table row is “Orijinal Jamba2 (3B)”, distinct from the Turkish SFT and CPT rows. The source does not identify an immutable upstream revision for the original Jamba2 baseline. Do not transfer these scores to the Turkish fine-tunes.
Source serda-dev, Jamba2-3B-Turkish README benchmark comparison
GSM8K-TR seeded 200-question subset, short-prompt accuracy (Show measurement, test conditions, and source)
Source value
19
Metric
accuracy
Unit
%
Benchmark version
malhajar/gsm8k-tr, 200-question sample, seed 20261014
Category
mathematics
Direction
Higher is better
Sample
200
Source type
community-reported
Evaluator
serda-dev, Jamba2-3B-Turkish README benchmark comparison
Status
active
Retrieved at
2026-10-20
Methodology
Community-reported result. The README states these accuracy results were obtained 2026-10-14 under the same protocol. The measured row is “Orijinal Jamba2 (3B)”. The upstream immutable revision for that baseline is not stated. Dataset: malhajar/gsm8k-tr; the same 200 questions selected with seed 20261014; short prompt.
Context
The exact table row is “Orijinal Jamba2 (3B)”, distinct from the Turkish SFT and CPT rows. The source does not identify an immutable upstream revision for the original Jamba2 baseline. Do not transfer these scores to the Turkish fine-tunes.
Source serda-dev, Jamba2-3B-Turkish README benchmark comparison
GSM8K-TR seeded 200-question subset, eight-shot accuracy (Show measurement, test conditions, and source)
Source value
20
Metric
accuracy
Unit
%
Benchmark version
malhajar/gsm8k-tr, 200-question sample, seed 20261014
Category
mathematics
Direction
Higher is better
Sample
200
Source type
community-reported
Evaluator
serda-dev, Jamba2-3B-Turkish README benchmark comparison
Status
active
Retrieved at
2026-10-20
Methodology
Community-reported result. The README states these accuracy results were obtained 2026-10-14 under the same protocol. The measured row is “Orijinal Jamba2 (3B)”. The upstream immutable revision for that baseline is not stated. Dataset: malhajar/gsm8k-tr; the same 200 questions selected with seed 20261014; eight fixed GSM8K-CoT examples translated into Turkish.
Context
The exact table row is “Orijinal Jamba2 (3B)”, distinct from the Turkish SFT and CPT rows. The source does not identify an immutable upstream revision for the original Jamba2 baseline. Do not transfer these scores to the Turkish fine-tunes.
Source serda-dev, Jamba2-3B-Turkish README benchmark comparison
IFEval-tr strict instruction-following subset (Show measurement, test conditions, and source)
Source value
46.7
Metric
Strict pass rate, requiring all specified conditions to be satisfied
Unit
%
Direction
Higher is better
Sample
120
Source type
community-reported
Evaluator
serda-dev, Jamba2-3B-Turkish evaluation report
Status
active
Retrieved at
2026-10-23
Methodology
Community-reported result. The report identifies the original checkpoint as ai21labs/AI21-Jamba2-3B at revision 525c6c8e1d9f5bddedfbdc1dbb0ade2df84230c9. It evaluates 120 deterministic Turkish-compatible examples from bezir/IFEval_tr, requires all conditions for a strict pass, uses greedy BF16 decoding with a 2,048-token output cap, and dates the run 2026-10-14.
Context
This is a small Turkish-compatible subset, not the full benchmark or a factual-accuracy score. The linked JSON containing Wilson 95% intervals was unavailable, so no interval is reported.
Source serda-dev, Jamba2-3B-Turkish evaluation report
IFBench (Show measurement, test conditions, and source)
Source value
0.36
Metric
IFBench
Unit
ratio
Category
source-specific
Direction
Higher is better
Source type
vendor-reported
Evaluator
AI21 announcement/HF chart
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. IFBench, Collie, IFEval and FACTS; Enterprise Reliability Score is their simple four-score average. Separate 100-prompt blind human side-by-side against Ministral3 14B, counterbalanced order.
Context
Source-specific observation; it is not a shared comparison cohort. AI21 self-reported comparison, not an external evaluator record
Source AI21 announcement/HF chart
Collie (Show measurement, test conditions, and source)
Source value
0.24
Metric
Collie
Unit
ratio
Category
source-specific
Direction
Higher is better
Source type
vendor-reported
Evaluator
AI21 announcement/HF chart
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. IFBench, Collie, IFEval and FACTS; Enterprise Reliability Score is their simple four-score average. Separate 100-prompt blind human side-by-side against Ministral3 14B, counterbalanced order.
Context
Source-specific observation; it is not a shared comparison cohort. AI21 self-reported comparison, not an external evaluator record
Source AI21 announcement/HF chart
IFEval (Show measurement, test conditions, and source)
Source value
0.93
Metric
IFEval
Unit
ratio
Category
source-specific
Direction
Higher is better
Source type
vendor-reported
Evaluator
AI21 announcement/HF chart
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. IFBench, Collie, IFEval and FACTS; Enterprise Reliability Score is their simple four-score average. Separate 100-prompt blind human side-by-side against Ministral3 14B, counterbalanced order.
Context
Source-specific observation; it is not a shared comparison cohort. AI21 self-reported comparison, not an external evaluator record
Source AI21 announcement/HF chart
FACTS (Show measurement, test conditions, and source)
Source value
0.54
Metric
FACTS
Unit
ratio
Category
source-specific
Direction
Higher is better
Source type
vendor-reported
Evaluator
AI21 announcement/HF chart
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. IFBench, Collie, IFEval and FACTS; Enterprise Reliability Score is their simple four-score average. Separate 100-prompt blind human side-by-side against Ministral3 14B, counterbalanced order.
Context
Source-specific observation; it is not a shared comparison cohort. AI21 self-reported comparison, not an external evaluator record
Source AI21 announcement/HF chart
Enterprise Reliability Score (average) (Show measurement, test conditions, and source)
Source value
0.52
Metric
Enterprise Reliability Score (average)
Unit
ratio
Category
source-specific
Direction
Higher is better
Source type
vendor-reported
Evaluator
AI21 announcement/HF chart
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. IFBench, Collie, IFEval and FACTS; Enterprise Reliability Score is their simple four-score average. Separate 100-prompt blind human side-by-side against Ministral3 14B, counterbalanced order.
Context
Source-specific observation; it is not a shared comparison cohort. AI21 self-reported comparison, not an external evaluator record
Source AI21 announcement/HF chart

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

The 100 most popular models of this category, 0 to 100, higher is more popular

20 of 101 models

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 20, 2026

Jamba2-3BDevstral Small 34Phi-4-multimodal-instruct 34Gemma 3n E2B IT 34Command A 33

Method changes in the timeline, a jump at these points is no market event:

  1. 1November 1, 2026: Value only with search or usage data, 30-day attention
  2. 2November 1, 2026: Method version 4 (common scale, shrinkage)
  3. 3November 9, 2026: New topics measurement V20 (daily quorum, short forms)

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

All specifications

Input and output

Context window
256K
Source: Context window

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Release
January 8, 2026
Source: Release
License
Apache 2.0
Source: License
Availability
AI21 Studio and Hugging Face
Source: Availability
Shutdown
None announced
no source

Frequently asked questions about Jamba2-3B

Can I run Jamba2-3B locally?

Yes, the weights are open.

Stated by AI21

Open weights
yes
License
Apache 2.0
Parameters
3B
active per token
n/a

Sources and data date

Show all 10 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.