Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Official1
huggingface.co2025-07-08officialauthorityHN front pageabout the SmolLM3 familyHN 388 pts · 79 comments3,203 words
SmolLM3 is a fully open 3B parameter model trained on 11T tokens that outperforms Llama-3.2-3B and Qwen2.5-3B while staying competitive with larger 4B alternatives. The model utilizes Grouped Query Attention with 4 groups and was trained on 384 H100 GPUs for 24 days using a global batch size of 2.36M tokens. It supports multilingual input for six languages and handles long contexts up to 128k tokens via NoPE and YaRN extrapolation.written, reviewed, cited
- The model removes weight decay from embedding layers to improve training stability.
- Users activate reasoning or non-reasoning modes by including the /think and /no_think flags in the system prompt.
Analysis4
github.comauthority858 words
An independent benchmark evaluates HuggingFace's SmolLM3-3B base model, a 3.08B dense decoder with GQA and NoPE trained on 11.2T tokens, using standardized vLLM tools on rented NVIDIA A40 hardware. The Apache-2.0 licensed model achieves 35.35% on GPQA Diamond via loglikelihood scoring, while generative tasks like MMLU-Pro show lower scores due to format penalties inherent to base checkpoints. Results match vendor figures for GSM8K and HumanEval+ within standard error margins.written, reviewed, cited
- GPQA Diamond loglikelihood scoring avoids instruction-following penalties, yielding 35.35% close to vendor instruct figures.
- GPQA Diamond evaluation requires enforce_eager=True to prevent out-of-memory errors on standard 48 GB GPUs.
- AIME 2024+2025 combined score of 1.7% reflects honest base-model performance without instruction tuning.
aial.ie1,282 words
HuggingFace released SmolLM3-3B, a fine-tuned model published on 2025-07-08. Its public summary scored 82.50% for Transparency and 86.01% for Usefulness, receiving a B+ grade in both categories. The document scored particularly well on disclosing public data sources, a key requirement under Article 53(1)(d), though it lacked information about other AI models used to generate synthetic data. The model dependencies should include SmolLM3-3B-Base.written, reviewed, cited
bestllmfor.com244 words
Running SmolLM3 3B locally with Q4 quantization requires about 2 GB of VRAM. The model generates tokens at a rate where beyond 20 t/s reading is comfortable.generated, cited
smollm3.orgabout the SmolLM3 family1,745 words
SmolLM3 is a 3 billion parameter language model available in transformers v4.53.0 or later that delivers exceptional performance in reasoning, multilingual understanding, and long context processing. The model supports extended thinking modes with reasoning traces and tool calling via xml_tools or python_tools, recommending temperature=0.6 and top_p=0.95 for optimal results. It can generate outputs up to 32768 new tokens while maintaining efficiency for edge computing scenarios and resource-constrained deployments.written, reviewed, cited
- SmolLM3 supports tool calling using xml_tools for standard functions or python_tools for Python function calls.
- The model features an extended thinking mode controlled via system prompts with /think and /no_think flags.
- SmolLM3 is engineered to deliver comparable performance with significantly reduced resource requirements compared to larger models.
