Skip to main content

AI21Language modelOpen source

Jamba Reasoning 3B

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

PopularityRank 277 of 342 language models

25out of 100new calculation

based on two data groups, mostly search, range 14 to 36

Summary

Jamba Reasoning 3B pairs reasoning with a memory-efficient SSM-Transformer architecture. AI21 reports a KV cache eight times smaller than that of a conventional Transformer. At a 32K-token context, the provider measures 40 output tokens per second on an M3 MacBook Pro.

AI21 releases the 3B model under Apache 2.0 with a 256K context window. The provider documents local use through LM Studio and llama.cpp, among other options.

All benchmark results

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

Show all 28 published benchmark results

Page 1 of 3

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.

CritPt v2: Overall (Show measurement, test conditions, and source)
Source value
0
Metric
pass@1 accuracy
Unit
%
Benchmark version
v2
Category
reasoning
Task
overall
Task label
Overall
Direction
Higher is better
Source type
independent-evaluator
Evaluator
Artificial Analysis
Status
disputed
Retrieved at
2026-11-02
Methodology
Exact AA configuration: Jamba Reasoning 3B (release jamba-reasoning-3b, row jamba-reasoning-3b, UUID f78138d6-2e04-4a84-919a-20d177cb6ff1, reasoning=true, effort=not specified, suffix=not specified). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The full CritPt v2 run covers 70 questions with five repeats. The official grading server reports pass@1.. Source model ID: artificial-analysis/f78138d6-2e04-4a84-919a-20d177cb6ff1/jamba-reasoning-3b/jamba-reasoning-3b. Source model name: Jamba Reasoning 3B. Provider reported by source: AI21 Labs. Model snapshot: Jamba Reasoning 3B. Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Tool policy: Python functions, symbolic expressions, and numerical answers; official grading server. Attempts: 70 questions, five repeats per question, pass@1. Reasoning: reasoning enabled. Score display: 0.0%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-11-02-67d46be20a5b. Source payload hash: 67d46be20a5b412db6cfb60ad190441fcaa5513500b8fd47d033168d0e97b6b2. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-11-02. Source cohort: artificial-analysis-critpt-v2:v2:standalone:reasoning:effort=unspecified:suffix=unspecified
Comparison methodology
Exact AA configuration: Jamba Reasoning 3B (release jamba-reasoning-3b, row jamba-reasoning-3b, UUID f78138d6-2e04-4a84-919a-20d177cb6ff1, reasoning=true, effort=not specified, suffix=not specified). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. The full CritPt v2 run covers 70 questions with five repeats. The official grading server reports pass@1.. Tool policy: Python functions, symbolic expressions, and numerical answers; official grading server. Attempts: 70 questions, five repeats per question, pass@1. Reasoning: reasoning enabled
Context
This AA source field contains the original CritPt score. The earlier CritPt v2 attribution was incorrect; the value and former label remain as disputed historical evidence.. The source dataset marks this result as disputed. It remains visible as documented evidence, but is not reliable comparative proof.
Source Artificial Analysis
IFBench: Overall (Show measurement, test conditions, and source)
Source value
52.4489795918
Metric
IFBench prompt-level loose accuracy (pass@1)
Unit
%
Benchmark version
AA test set: 294 questions
Category
general
Task
overall
Task label
Overall
Direction
Higher is better
Source type
independent-evaluator
Evaluator
Artificial Analysis
Status
superseded
Retrieved at
2026-11-02
Methodology
Exact AA configuration: Jamba Reasoning 3B (release jamba-reasoning-3b, row jamba-reasoning-3b, UUID f78138d6-2e04-4a84-919a-20d177cb6ff1, reasoning=true, effort=not specified, suffix=not specified). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. Artificial Analysis runs 294 questions from the IFBench test set with five repeats per question. The official evaluator scores prompt-level accuracy in loose mode as pass@1.. Source model ID: artificial-analysis/f78138d6-2e04-4a84-919a-20d177cb6ff1/jamba-reasoning-3b/jamba-reasoning-3b. Source model name: Jamba Reasoning 3B. Provider reported by source: AI21 Labs. Model snapshot: Jamba Reasoning 3B. Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Tool policy: Single-turn answer evaluated with the official IFBench source code. Attempts: pass@1, 5 repeats per question. Reasoning: reasoning enabled. Score display: 52.4%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-11-02-67d46be20a5b. Source payload hash: 67d46be20a5b412db6cfb60ad190441fcaa5513500b8fd47d033168d0e97b6b2. Parser version: artificial-analysis-rsc-models-v3. Imported at: 2026-11-02. Source cohort: artificial-analysis-ifbench:AA test set: 294 questions:standalone:reasoning:effort=unspecified:suffix=unspecified
Comparison methodology
Exact AA configuration: Jamba Reasoning 3B (release jamba-reasoning-3b, row jamba-reasoning-3b, UUID f78138d6-2e04-4a84-919a-20d177cb6ff1, reasoning=true, effort=not specified, suffix=not specified). AA source value for average reasoning tokens before the answer: 2000. Its methodology averages 60 varied prompts and assumes 2,000 when no average is available. This is not a token cap. Artificial Analysis runs 294 questions from the IFBench test set with five repeats per question. The official evaluator scores prompt-level accuracy in loose mode as pass@1.. Tool policy: Single-turn answer evaluated with the official IFBench source code. Attempts: pass@1, 5 repeats per question. Reasoning: reasoning enabled
Context
The often cited 58 refers to verifiable out-of-domain constraint types, not the number of questions in AA's test run.. This result is retained as historical evidence and is not interchangeable with current results.
Source Artificial Analysis
IFBench: Overall (Show measurement, test conditions, and source)
Source value
52.4489795918
Metric
IFBench prompt-level loose accuracy (pass@1)
Unit
%
Benchmark version
AA test set: 294 questions
Category
general
Task
overall
Task label
Overall
Direction
Higher is better
Source type
independent-evaluator
Evaluator
Artificial Analysis
Status
active
Retrieved at
2026-11-07
Methodology
Exact AA configuration: Jamba Reasoning 3B (release jamba-reasoning-3b, row jamba-reasoning-3b, UUID f78138d6-2e04-4a84-919a-20d177cb6ff1, reasoning=true, effort=not specified, suffix=not specified). This source capture does not provide an average reasoning-token count. Artificial Analysis runs 294 questions from the IFBench test set with five repeats per question. The official evaluator scores prompt-level accuracy in loose mode as pass@1.. Source model ID: artificial-analysis/f78138d6-2e04-4a84-919a-20d177cb6ff1/jamba-reasoning-3b/jamba-reasoning-3b. Source model name: Jamba Reasoning 3B. Provider reported by source: AI21 Labs. Model snapshot: Jamba Reasoning 3B. Evaluator: Artificial Analysis. Dataset type: artificial-analysis-evaluation. Tool policy: Single-turn answer evaluated with the official IFBench source code. Attempts: pass@1, 5 repeats per question. Reasoning: reasoning enabled. Score display: 52.4%. Source ID: artificial-analysis. Evidence capture ID: aa-model-page-2026-11-07-b96e3ffc8c5b. Source payload hash: b96e3ffc8c5b052aa56f96d05e89d02b8cc234a75085820fca0ca2a39e498274. Parser version: artificial-analysis-rsc-models-v4. Imported at: 2026-11-07. Source cohort: artificial-analysis-ifbench:AA test set: 294 questions:standalone:reasoning:effort=unspecified:suffix=unspecified
Comparison methodology
Exact AA configuration: Jamba Reasoning 3B (release jamba-reasoning-3b, row jamba-reasoning-3b, UUID f78138d6-2e04-4a84-919a-20d177cb6ff1, reasoning=true, effort=not specified, suffix=not specified). This source capture does not provide an average reasoning-token count. Artificial Analysis runs 294 questions from the IFBench test set with five repeats per question. The official evaluator scores prompt-level accuracy in loose mode as pass@1.. Tool policy: Single-turn answer evaluated with the official IFBench source code. Attempts: pass@1, 5 repeats per question. Reasoning: reasoning enabled
Context
The often cited 58 refers to verifiable out-of-domain constraint types, not the number of questions in AA's test run.
Source Artificial Analysis
IFBench (Show measurement, test conditions, and source)
Source value
52
Metric
IFBench
Unit
%
Category
source-specific
Direction
Higher is better
Source type
vendor-reported
Evaluator
AI21 announcement chart pixels
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Component chart against small-model peers; separate composite combines MMLU-Pro, LiveCodeBench, GPQA-Diamond, SciCode, IFBench and HLE at 32K context on Apple M3. Composite is not one benchmark.
Context
Source-specific observation; it is not a shared comparison cohort. AI21 self-reported chart; no exact independent record
Source AI21 announcement chart pixels
Humanity's Last Exam (Show measurement, test conditions, and source)
Source value
6
Metric
Humanity's Last Exam
Unit
%
Category
source-specific
Direction
Higher is better
Source type
vendor-reported
Evaluator
AI21 announcement chart pixels
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Component chart against small-model peers; separate composite combines MMLU-Pro, LiveCodeBench, GPQA-Diamond, SciCode, IFBench and HLE at 32K context on Apple M3. Composite is not one benchmark.
Context
Source-specific observation; it is not a shared comparison cohort. AI21 self-reported chart; no exact independent record
Source AI21 announcement chart pixels
MMLU-Pro (Show measurement, test conditions, and source)
Source value
61
Metric
MMLU-Pro
Unit
%
Category
source-specific
Direction
Higher is better
Source type
vendor-reported
Evaluator
AI21 announcement chart pixels
Status
active
Retrieved at
2026-10-04
Methodology
Vendor-reported result. Component chart against small-model peers; separate composite combines MMLU-Pro, LiveCodeBench, GPQA-Diamond, SciCode, IFBench and HLE at 32K context on Apple M3. Composite is not one benchmark.
Context
Source-specific observation; it is not a shared comparison cohort. AI21 self-reported chart; no exact independent record
Source AI21 announcement chart pixels
DEXBench CruxEval Rexec pass@1 (Show measurement, test conditions, and source)
Source value
0
Metric
pass@1
Unit
%
Benchmark version
DEXBench, ACL 2026 Table 1
Category
code-reasoning
Direction
Higher is better
Sample
298
Source type
independent-evaluator
Evaluator
DEXBench, The Path Not Taken, ACL 2026 Table 1
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact table row: Jamba Reasoning 3B. DEXBench evaluates forward execution reasoning (Rexec), counterfactual reasoning (Rcf), and joint dual-path reasoning (Rdual), each at pass@1 and pass@5 with one-shot prompts. Per-dataset N is recorded. The paper does not disclose an immutable checkpoint revision or full generation settings.
Context
The paper reports pass@k in percent and no confidence interval. The task counts refer to paired forward/backward program instances, not to a number of generated candidates.
Source DEXBench, The Path Not Taken, ACL 2026 Table 1
DEXBench CruxEval Rexec pass@5 (Show measurement, test conditions, and source)
Source value
0
Metric
pass@5
Unit
%
Benchmark version
DEXBench, ACL 2026 Table 1
Category
code-reasoning
Direction
Higher is better
Sample
298
Source type
independent-evaluator
Evaluator
DEXBench, The Path Not Taken, ACL 2026 Table 1
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact table row: Jamba Reasoning 3B. DEXBench evaluates forward execution reasoning (Rexec), counterfactual reasoning (Rcf), and joint dual-path reasoning (Rdual), each at pass@1 and pass@5 with one-shot prompts. Per-dataset N is recorded. The paper does not disclose an immutable checkpoint revision or full generation settings.
Context
The paper reports pass@k in percent and no confidence interval. The task counts refer to paired forward/backward program instances, not to a number of generated candidates.
Source DEXBench, The Path Not Taken, ACL 2026 Table 1
DEXBench CruxEval Rcf pass@1 (Show measurement, test conditions, and source)
Source value
9.4
Metric
pass@1
Unit
%
Benchmark version
DEXBench, ACL 2026 Table 1
Category
code-reasoning
Direction
Higher is better
Sample
298
Source type
independent-evaluator
Evaluator
DEXBench, The Path Not Taken, ACL 2026 Table 1
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact table row: Jamba Reasoning 3B. DEXBench evaluates forward execution reasoning (Rexec), counterfactual reasoning (Rcf), and joint dual-path reasoning (Rdual), each at pass@1 and pass@5 with one-shot prompts. Per-dataset N is recorded. The paper does not disclose an immutable checkpoint revision or full generation settings.
Context
The paper reports pass@k in percent and no confidence interval. The task counts refer to paired forward/backward program instances, not to a number of generated candidates.
Source DEXBench, The Path Not Taken, ACL 2026 Table 1
DEXBench CruxEval Rcf pass@5 (Show measurement, test conditions, and source)
Source value
14.8
Metric
pass@5
Unit
%
Benchmark version
DEXBench, ACL 2026 Table 1
Category
code-reasoning
Direction
Higher is better
Sample
298
Source type
independent-evaluator
Evaluator
DEXBench, The Path Not Taken, ACL 2026 Table 1
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact table row: Jamba Reasoning 3B. DEXBench evaluates forward execution reasoning (Rexec), counterfactual reasoning (Rcf), and joint dual-path reasoning (Rdual), each at pass@1 and pass@5 with one-shot prompts. Per-dataset N is recorded. The paper does not disclose an immutable checkpoint revision or full generation settings.
Context
The paper reports pass@k in percent and no confidence interval. The task counts refer to paired forward/backward program instances, not to a number of generated candidates.
Source DEXBench, The Path Not Taken, ACL 2026 Table 1
DEXBench CruxEval Rdual pass@1 (Show measurement, test conditions, and source)
Source value
0
Metric
pass@1
Unit
%
Benchmark version
DEXBench, ACL 2026 Table 1
Category
code-reasoning
Direction
Higher is better
Sample
298
Source type
independent-evaluator
Evaluator
DEXBench, The Path Not Taken, ACL 2026 Table 1
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact table row: Jamba Reasoning 3B. DEXBench evaluates forward execution reasoning (Rexec), counterfactual reasoning (Rcf), and joint dual-path reasoning (Rdual), each at pass@1 and pass@5 with one-shot prompts. Per-dataset N is recorded. The paper does not disclose an immutable checkpoint revision or full generation settings.
Context
The paper reports pass@k in percent and no confidence interval. The task counts refer to paired forward/backward program instances, not to a number of generated candidates.
Source DEXBench, The Path Not Taken, ACL 2026 Table 1
DEXBench CruxEval Rdual pass@5 (Show measurement, test conditions, and source)
Source value
0
Metric
pass@5
Unit
%
Benchmark version
DEXBench, ACL 2026 Table 1
Category
code-reasoning
Direction
Higher is better
Sample
298
Source type
independent-evaluator
Evaluator
DEXBench, The Path Not Taken, ACL 2026 Table 1
Status
active
Retrieved at
2026-10-17
Methodology
Independent evaluation. Exact table row: Jamba Reasoning 3B. DEXBench evaluates forward execution reasoning (Rexec), counterfactual reasoning (Rcf), and joint dual-path reasoning (Rdual), each at pass@1 and pass@5 with one-shot prompts. Per-dataset N is recorded. The paper does not disclose an immutable checkpoint revision or full generation settings.
Context
The paper reports pass@k in percent and no confidence interval. The task counts refer to paired forward/backward program instances, not to a number of generated candidates.
Source DEXBench, The Path Not Taken, ACL 2026 Table 1

Page 1 of 3

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

Jamba Reasoning 3Bo1 25Baichuan M2 32B 25Amazon Nova Premier 25Kimi K2.7 Code Highspeed 25

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, with processing up to 1M tokens described
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
October 8, 2025
Source: Release
License
Apache 2.0
Source: License
Shutdown
None announced
no source

Frequently asked questions about Jamba Reasoning 3B

Can I run Jamba Reasoning 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 12 sources

Benchmark sources at a glance

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

EvaluatorTest seriesComposite indicesRecorded active measurements
Artificial Analysis101

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

TypeEvidence and data date
Research statusResearch date October 4, 2026. 4 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not found in the inspected sources

benchmark-weights-revision

Note

The paper identifies the Hugging Face repository but does not bind the evaluation to a commit SHA or immutable weight revision.

Additional sourceAI21 Jamba Reasoning 3B announcement (retrieved October 3, 2026; October 4, 2026; October 19, 2026) · AI21 Jamba Reasoning 3B announcement · Editorial description reviewed October 4, 2026
Additional sourceArtificial Analysis (retrieved November 2, 2026)
Additional sourceArtificial Analysis (retrieved November 2, 2026; November 7, 2026)
Additional sourceDirected Attention is All You Need: Profiling Style from Limited Text Data (SIGTURK 2026) (retrieved October 16, 2026)
Additional sourceDEXBench, The Path Not Taken, ACL 2026 Table 1 (retrieved October 17, 2026)
Additional sourceModel metadata (retrieved October 16, 2026)
Additional sourceOpenRouter: Popularity index (derived), CC BY 4.0
Additional sourceVercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional sourceOllama: cumulative downloads for the popularity index
Additional sourceReplicate: cumulative runs for the popularity index
Additional sourceGoogle Trending Now: limited feed selection for Germany and the US
Additional sourceCreative Commons Attribution 4.0