TIILanguage model
Falcon 3 10B
Falcon 3 10B has no overall index value yet because there are too few independent benchmark results.
PopularityRank 620 of 785
26out of 100-2 vs. last month
based on search, usage and attention
Summary
Falcon 3 10B is TII’s larger open model in the Falcon 3 cycle. TII describes it as the result of depth upscaling, where redundant layers from the Falcon 3 7B training line were duplicated and then continued with another 2 trillion high-quality tokens.
The base checkpoint has a 32,768-token context and uses the Falcon license. A separate instruction-tuned derivative is available for dialogue.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 27 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.
Visual generalization (average per-class accuracy): AWA2 (Show measurement, test conditions, and source)
- Source value
- 60
- Score
- 60
- Metric
- Average per-class classification accuracy on unaligned classes
- Unit
- %
- Benchmark version
- TMLR 01/2026, Appendix Table 12
- Category
- visual classification generalization
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Better Language Models Exhibit Higher Visual Alignment (TMLR, 01/2026; author-hosted full paper)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Falcon3-10B-Base is used as a frozen decoder-based text encoder in the paper's CLIP-like visual-generalization framework: last-token-pooled text features with DINOv2 ViT-L/14 image features and a learned linear projection. Evaluation reports average per-class classification accuracy on unaligned/unseen classes. The paper says its results average five initialization seeds; Table 12 reports the raw cells AWA2: 60.0 ± 6.8; CUB: 44.6 ± 1.5; Aircraft: 32.7 ± 2.8; ImageNet+: 34.6 ± 0.4; Average: 43.0, but does not define the statistic after ±. The scalar score stores only the reported central value; ± is preserved in evidence and is not interpreted as a CI or standard deviation. No exact weight hash or checkpoint revision is given.
- Context
- This is a result for the combined visual encoder/projection evaluation, not standalone LLM generation or intrinsic language-model accuracy. The paper does not define the reported ± statistic; it is not converted into a CI.
Visual generalization (average per-class accuracy): CUB (Show measurement, test conditions, and source)
- Source value
- 44.6
- Score
- 44.6
- Metric
- Average per-class classification accuracy on unaligned classes
- Unit
- %
- Benchmark version
- TMLR 01/2026, Appendix Table 12
- Category
- visual classification generalization
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Better Language Models Exhibit Higher Visual Alignment (TMLR, 01/2026; author-hosted full paper)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Falcon3-10B-Base is used as a frozen decoder-based text encoder in the paper's CLIP-like visual-generalization framework: last-token-pooled text features with DINOv2 ViT-L/14 image features and a learned linear projection. Evaluation reports average per-class classification accuracy on unaligned/unseen classes. The paper says its results average five initialization seeds; Table 12 reports the raw cells AWA2: 60.0 ± 6.8; CUB: 44.6 ± 1.5; Aircraft: 32.7 ± 2.8; ImageNet+: 34.6 ± 0.4; Average: 43.0, but does not define the statistic after ±. The scalar score stores only the reported central value; ± is preserved in evidence and is not interpreted as a CI or standard deviation. No exact weight hash or checkpoint revision is given.
- Context
- This is a result for the combined visual encoder/projection evaluation, not standalone LLM generation or intrinsic language-model accuracy. The paper does not define the reported ± statistic; it is not converted into a CI.
Visual generalization (average per-class accuracy): Aircraft (Show measurement, test conditions, and source)
- Source value
- 32.7
- Score
- 32.7
- Metric
- Average per-class classification accuracy on unaligned classes
- Unit
- %
- Benchmark version
- TMLR 01/2026, Appendix Table 12
- Category
- visual classification generalization
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Better Language Models Exhibit Higher Visual Alignment (TMLR, 01/2026; author-hosted full paper)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Falcon3-10B-Base is used as a frozen decoder-based text encoder in the paper's CLIP-like visual-generalization framework: last-token-pooled text features with DINOv2 ViT-L/14 image features and a learned linear projection. Evaluation reports average per-class classification accuracy on unaligned/unseen classes. The paper says its results average five initialization seeds; Table 12 reports the raw cells AWA2: 60.0 ± 6.8; CUB: 44.6 ± 1.5; Aircraft: 32.7 ± 2.8; ImageNet+: 34.6 ± 0.4; Average: 43.0, but does not define the statistic after ±. The scalar score stores only the reported central value; ± is preserved in evidence and is not interpreted as a CI or standard deviation. No exact weight hash or checkpoint revision is given.
- Context
- This is a result for the combined visual encoder/projection evaluation, not standalone LLM generation or intrinsic language-model accuracy. The paper does not define the reported ± statistic; it is not converted into a CI.
Visual generalization (average per-class accuracy): ImageNet+ (Show measurement, test conditions, and source)
- Source value
- 34.6
- Score
- 34.6
- Metric
- Average per-class classification accuracy on unaligned classes
- Unit
- %
- Benchmark version
- TMLR 01/2026, Appendix Table 12
- Category
- visual classification generalization
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Better Language Models Exhibit Higher Visual Alignment (TMLR, 01/2026; author-hosted full paper)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Falcon3-10B-Base is used as a frozen decoder-based text encoder in the paper's CLIP-like visual-generalization framework: last-token-pooled text features with DINOv2 ViT-L/14 image features and a learned linear projection. Evaluation reports average per-class classification accuracy on unaligned/unseen classes. The paper says its results average five initialization seeds; Table 12 reports the raw cells AWA2: 60.0 ± 6.8; CUB: 44.6 ± 1.5; Aircraft: 32.7 ± 2.8; ImageNet+: 34.6 ± 0.4; Average: 43.0, but does not define the statistic after ±. The scalar score stores only the reported central value; ± is preserved in evidence and is not interpreted as a CI or standard deviation. No exact weight hash or checkpoint revision is given.
- Context
- This is a result for the combined visual encoder/projection evaluation, not standalone LLM generation or intrinsic language-model accuracy. The paper does not define the reported ± statistic; it is not converted into a CI.
Visual generalization (average per-class accuracy): Average (Show measurement, test conditions, and source)
- Source value
- 43
- Score
- 43
- Metric
- Average per-class classification accuracy on unaligned classes
- Unit
- %
- Benchmark version
- TMLR 01/2026, Appendix Table 12
- Category
- visual classification generalization
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Better Language Models Exhibit Higher Visual Alignment (TMLR, 01/2026; author-hosted full paper)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Falcon3-10B-Base is used as a frozen decoder-based text encoder in the paper's CLIP-like visual-generalization framework: last-token-pooled text features with DINOv2 ViT-L/14 image features and a learned linear projection. Evaluation reports average per-class classification accuracy on unaligned/unseen classes. The paper says its results average five initialization seeds; Table 12 reports the raw cells AWA2: 60.0 ± 6.8; CUB: 44.6 ± 1.5; Aircraft: 32.7 ± 2.8; ImageNet+: 34.6 ± 0.4; Average: 43.0, but does not define the statistic after ±. The scalar score stores only the reported central value; ± is preserved in evidence and is not interpreted as a CI or standard deviation. No exact weight hash or checkpoint revision is given.
- Context
- This is a result for the combined visual encoder/projection evaluation, not standalone LLM generation or intrinsic language-model accuracy. The paper does not define the reported ± statistic; it is not converted into a CI.
MMLU (Show measurement, test conditions, and source)
- Source value
- 73.1
- Score
- 73.1
- Metric
- raw score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Technology Innovation Institute Falcon3-10B-Base model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Falcon3-10B-Base column in TII internal pipeline table. Source states lm-evaluation-harness, raw scores, and same batch size across compared models; numeric unit/scale and task snapshot dates are not stated.; 5-shot.
- Context
- The source gives a numeric score but does not label its unit or scale; the raw value is retained without converting it to percent.
MMLU-PRO (Show measurement, test conditions, and source)
- Source value
- 42.5
- Score
- 42.5
- Metric
- raw score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Technology Innovation Institute Falcon3-10B-Base model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Falcon3-10B-Base column in TII internal pipeline table. Source states lm-evaluation-harness, raw scores, and same batch size across compared models; numeric unit/scale and task snapshot dates are not stated.; 5-shot.
- Context
- The source gives a numeric score but does not label its unit or scale; the raw value is retained without converting it to percent.
IFEval (Show measurement, test conditions, and source)
- Source value
- 36.4
- Score
- 36.4
- Metric
- raw score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Technology Innovation Institute Falcon3-10B-Base model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Falcon3-10B-Base column in TII internal pipeline table. Source states lm-evaluation-harness, raw scores, and same batch size across compared models; numeric unit/scale and task snapshot dates are not stated.; shot count not stated.
- Context
- The source gives a numeric score but does not label its unit or scale; the raw value is retained without converting it to percent.
GSM8K (Show measurement, test conditions, and source)
- Source value
- 81.4
- Score
- 81.4
- Metric
- raw score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Technology Innovation Institute Falcon3-10B-Base model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Falcon3-10B-Base column in TII internal pipeline table. Source states lm-evaluation-harness, raw scores, and same batch size across compared models; numeric unit/scale and task snapshot dates are not stated.; 5-shot.
- Context
- The source gives a numeric score but does not label its unit or scale; the raw value is retained without converting it to percent.
MATH Lvl-5 (Show measurement, test conditions, and source)
- Source value
- 22.9
- Score
- 22.9
- Metric
- raw score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Technology Innovation Institute Falcon3-10B-Base model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Falcon3-10B-Base column in TII internal pipeline table. Source states lm-evaluation-harness, raw scores, and same batch size across compared models; numeric unit/scale and task snapshot dates are not stated.; 4-shot.
- Context
- The source gives a numeric score but does not label its unit or scale; the raw value is retained without converting it to percent.
ARC Challenge (Show measurement, test conditions, and source)
- Source value
- 66.8
- Score
- 66.8
- Metric
- raw score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Technology Innovation Institute Falcon3-10B-Base model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Falcon3-10B-Base column in TII internal pipeline table. Source states lm-evaluation-harness, raw scores, and same batch size across compared models; numeric unit/scale and task snapshot dates are not stated.; 25-shot.
- Context
- The source gives a numeric score but does not label its unit or scale; the raw value is retained without converting it to percent.
GPQA (Show measurement, test conditions, and source)
- Source value
- 34.1
- Score
- 34.1
- Metric
- raw score
- Unit
- No unit provided
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Technology Innovation Institute Falcon3-10B-Base model card
- Status
- active
- Retrieved at
- 2026-10-15
- Methodology
- Vendor-reported result. Exact Falcon3-10B-Base column in TII internal pipeline table. Source states lm-evaluation-harness, raw scores, and same batch size across compared models; numeric unit/scale and task snapshot dates are not stated.; 0-shot.
- Context
- The source gives a numeric score but does not label its unit or scale; the raw value is retained without converting it to percent.
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 7, 2026
Falcon 3 10BCommand A 26Grok 4.20 Multi-Agent 26Llama 3.1 Nemotron Ultra 253B 26Aion 3.5 Mini 26
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as Falcon 3 10B, chosen by model line, developer, size, price and release date.
All specifications
Architecture
Input and output
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
Frequently asked questions about Falcon 3 10B
How large is the context window of Falcon 3 10B?
32,768 tokens. That is how much text Falcon 3 10B reads at once, about 50 A4 pages.
Page count as a rule of thumb from Artificial Analysis
GLM-5.1200K
DeepSeek-V3 0324164K
GLM-4.5131K
GPT OSS 120B131K
DeepSeek-V3.2128K
DeepSeek-R1128K
Falcon 3 10BRank 103 of 10333K
Can I run Falcon 3 10B locally?
Yes, the weights are open.
Stated by TII
- Open weights
- yes
- License
- Falcon License
- Parameters
- 10B
- 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.
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. |
| Output | Documented modalities. |
| Research status | Research date October 15, 2026. 3 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | score-unit-and-task-date Note Card calls values raw scores but does not label unit/scale or task snapshot dates; values remain unknown-unit raw scores. |
| Not published | checkpoint revision Note The paper names the exact Falcon3 base size/checkpoint family but does not give an immutable weight hash or revision. |
| Not published | sample counts Note No per-dataset scored example count is stated with Table 12; no sampleCount is inferred. |
| Unresolved | reported ± statistic Note Table 12 prints ± values and the paper states results average five initialization seeds, but does not define the statistic represented by ±. The raw text is retained in evidence and no CI is encoded. |
| Unresolved | Open LLM Leaderboard result configuration Note The exact card identifies the Open LLM Leaderboard and provides per-task metric labels and shot counts for six rows, but not a leaderboard snapshot/version, result date, sample sizes, confidence intervals, unit/scale, or score direction. The same card's TII internal-pipeline table gives different values for some same-named tasks; these remain separate evaluator records. |
| Unresolved | Open LLM Leaderboard average composite definition Note The exact card prints “Avg. | 27.59”, but does not state the composite's task set, weighting, normalization, sample count, direction, or formula. The source-reported value is not recomputed or treated as a rank. |
| Additional source | Technology Innovation Institute Falcon3-10B-Base model card (retrieved October 15, 2026; October 19, 2026) |
| Additional source | Better Language Models Exhibit Higher Visual Alignment (TMLR, 01/2026; author-hosted full paper) (retrieved October 17, 2026) |
| Additional source | Falcon 3 10B model card · Editorial description reviewed October 2, 2026 |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |