SmolLM3
Fully open 3B multilingual reasoner with a published training recipe
About
SmolLM3 is Hugging Face's 3-billion-parameter open language model, released in July 2025 under Apache 2.0 with the full training recipe published alongside the weights. It was trained on 11.2 trillion tokens, reasons in a dual mode that can be toggled with /think and /no_think flags, and supports six languages plus tool calling. At 3B it runs on consumer hardware without a dedicated GPU, though its reasoning depth and factual reliability trail much larger models.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Released 2025-07-08 by Hugging Face under Apache 2.0 with the full training recipe (data mixture, configs) published; 3.08B params, trained on 11.2T tokens; 64K trained context, 128K via YaRN; Dual-mode reasoning (/think, /no_think) and native tool calling verified from multiple writeups; Benchmark figures not collected this session; entry makes no benchmark claims. License: Apache-2.0 confirmed across sources..
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Released 2025-07-08 by Hugging Face under Apache 2.0 with the full training recipe (data mixture, configs) published
- · 3.08B params, trained on 11.2T tokens; 64K trained context, 128K via YaRN
- · Dual-mode reasoning (/think, /no_think) and native tool calling verified from multiple writeups
- · Benchmark figures not collected this session; entry makes no benchmark claims