Falcon

Name: Falcon
Author: Technology Innovation Institute

Technology Innovation InstituteOpen WeightsPending Human Review

Falcon is a TII open-weight language-model release. The initial Falcon family includes 7B and 40B models trained on RefinedWeb and curated corpora. Base and instruct checkpoints support completion and conversation. The family’s open commercial weight release is dated May 25; later removal of royalties and adoption of Apache licensing changed terms without creating a new checkpoint. Published license terms and checkpoint-specific hardware needs apply. Generated text can contain unsupported claims; sparse computation, reduced precision, or hybrid sequence modeling does not remove the need for task evaluation.

2023-05-25

7B, 40B

Dense decoder-only Transformer with multi-query attention

Apache-2.0 (current 7B/40B weights)

Compare with other models

Specifications

Parameters: 7B, 40B
Architecture: Dense decoder-only Transformer with multi-query attention
License: Apache-2.0 (current 7B/40B weights)
Context Window: 2,048 tokens
Type: text
Modalities: text

Benchmark Scores

Advanced Specifications

Model Family: Falcon
API Access: Not Available
Chat Interface: Not Available
Variants: 7B Base/Instruct40B Base/Instruct

Capabilities & Limitations

Capabilities: text generationinstruction following
Known Limitations: Capabilities vary across base and instruct checkpointsSmaller tiers have limited broad knowledge
Notable Use Cases: local language applicationsmodel researchenterprise customization

Resources

Related Models

Falcon-H1-Tiny

Technology Innovation Institute

Falcon-H1-Tiny is a TII open-weight language-model release. Falcon-H1-Tiny is a family of micro language models for targeted edge tasks. The release includes instruction, multilingual, code, tool-call, and reasoning checkpoints, including a 600M reasoning tier. These specializations trade broad knowledge for small memory and compute budgets. The official article source records January 15 publication. Published license terms and checkpoint-specific hardware needs apply. Generated text can contain unsupported claims; sparse computation, reduced precision, or hybrid sequence modeling does not remove the need for task evaluation.

Typetext

Parameters90M, 100M, 600M

2026-01-15

Open Weights

Details Compare

Falcon-H1-Arabic

Technology Innovation Institute

Falcon-H1-Arabic is a TII open-weight language-model release. Falcon-H1-Arabic specializes the hybrid Falcon architecture for Arabic, dialects, and bilingual enterprise work. Continued training and post-training focus on long-context instruction following and Arabic linguistic quality. The 3B tier supports 128K context, while 7B and 34B support 256K; the family-level context shown describes the larger tiers. Published license terms and checkpoint-specific hardware needs apply. Generated text can contain unsupported claims; sparse computation, reduced precision, or hybrid sequence modeling does not remove the need for task evaluation.

Typetext

Parameters3B, 7B, 34B

2026-01-05

Open Weights

Details Compare

Falcon-H1R

Technology Innovation Institute

Falcon-H1R is a TII open-weight language-model release. Falcon-H1R is the reasoning-specialized 7B hybrid model post-trained from Falcon-H1. It targets math, coding, and scientific reasoning with longer test-time computation and search options. Reported contest performance may use DeepConf or multiple candidates, and should not be interpreted as single-response accuracy. Published license terms and checkpoint-specific hardware needs apply. Generated text can contain unsupported claims; sparse computation, reduced precision, or hybrid sequence modeling does not remove the need for task evaluation.