Falcon3 3B
The UAE's pruned-and-healed 3B instruct model: 32K context, four languages, and official GGUF/GPTQ builds for local deployment.
About
Falcon3 3B is the mid-size member of the UAE Technology Innovation Institute's Falcon3 family, and it has an unusual origin story: it was pruned and healed from the 7B base model on 100 billion tokens rather than trained from scratch, using 1024 H100 GPUs. It is a 22-layer decoder with grouped-query attention, SwiGLU and RMSNorm, a 131K vocabulary, and official support for English, French, Spanish and Portuguese. The instruct version stretches context to 32K tokens.
TII's internal lm-evaluation-harness runs put the instruct model at 56.9 MMLU, 68.3 IFEval and 74.8 GSM8K, competitive with Llama-3.2-3B-Instruct and Qwen2.5-3B-Instruct on reasoning and instruction following, with particular strength on MATH and BBH. The license is the TII Falcon-LLM License 2.0, which is broadly permissive but carries acceptable-use conditions, so it is not a standard OSI license. Quantized GGUF and GPTQ builds ship officially, which makes it a practical pick for local deployment.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Falcon3-3B-Instruct (TII internal pipeline, lm-eval harness): MMLU 56.9, MMLU-Pro 29.7, IFEval 68.3, GSM8K 74.8, MATH Lvl-5 19.9, BBH 45.4, MT-Bench 7.2 (provider-reported)
- Peers on same harness: Llama-3.2-3B-Instruct (MMLU 61.2, GSM8K 76.8), Qwen2.5-3B-Instruct (MMLU 65.4, GSM8K 56.7) (provider-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: TII Falcon-LLM License 2.0, December 2024 release date, 32K instruct context and EN/FR/ES/PT language support all confirmed on the official tiiuae/Falcon3-3B-Instruct model card; benchmark table cross-checked across the Instruct, Base and GGUF model cards. Strengths: transparent internal benchmark table against named peers, official quantized builds, state-backed maintenance. Docked for: (1) benchmarks are TII's own pipeline with no independent replication found; (2) Falcon-LLM License 2.0 has acceptable-use terms, so it is not OSI-approved open source. License: TII Falcon-LLM License 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · TII UAE; released 2024-12
- · Pruned+healed from Falcon3-7B on 100B tokens
- · Internal lm-eval harness table; provider-reported