TII's first Arabic-specific model in the Falcon series, released May 21, 2025. Built on Falcon 3-7B as the base model and trained on 100 percent native Arabic datasets with no machine-translated content. The tokenizer was extended with 32,000 Arabic-specific tokens using a novel embedding initialization based on textual similarity. Training followed a multi-stage curriculum: general knowledge and dialect-rich Arabic content in early stages, followed by math, code, and reasoning in later phases. Post-training applied SFT with public and internally collected native Arabic instruction data, followed by DPO. Context window of 32,000 tokens. TII reports the model leads Arabic MMLU, Exams, MadinahQA, and AraTrust within its size range and publishes full evaluation details on Hugging Face.
| 信号 | 强度 | 权重 | 影响 |
|---|---|---|---|
| Pricingjust now | 100 | 25% | +25.0 |
| Capabilitiesjust now | 50 | 30% | +15.0 |
| Recencyjust now | 74 | 15% | +11.1 |
| Context Windowjust now | 72 | 15% | +10.7 |
| Output Capacityjust now | 65 | 15% | +9.8 |
把当前模型放回同一服务商最近的发布节奏中查看。
Falcon-H1-Arabic 34B Instruct
coding
Falcon-H1-Arabic 7B Instruct
coding
Falcon-H1-Arabic 3B Instruct
coding
Falcon Arabic 7B Instruct当前模型
coding
Falcon3 10B Instruct
coding
Falcon3 7B Instruct
coding
Falcon Mamba 7B Instruct
coding
社区和从业者反馈在基准测试和价格之上增加了真实世界的信号。
Share your experience with Falcon Arabic 7B Instruct and help the community make better decisions.