Small-model reasoning focus
The publisher positions InfiR as advancing reasoning while reducing adoption barriers through smaller model sizes.
Open Source Model Profile · InfiX-ai
InfiR-1B-Instruct
InfiR-1B-Instruct is a 1.50B-parameter reasoning model from InfiX-ai. Its model card documents chat-template usage, a full training recipe, and publisher-reported MMLU and GSM8K scores.
- Publisher
- InfiX-ai
- Task
- text-generation
- Model type
- llama
- License
- llama3.2
- Library
- Unknown
- Publication status
- Accepted · not indexed
Model overview
InfiR-1B-Instruct is published by InfiX-ai as an English text-generation model with LlamaForCausalLM architecture and a llama model type. Safetensors metadata reports 1,498,482,688 parameters, or about 1.50B. According to the model card, it targets reasoning in a small-model footprint, and captured metadata records a llama3.2 license.
Recorded capabilities
Published benchmark table
The card reports MMLU, GSM8K, MATH, HumanEval, and MBPP scores with same-scale comparison figures.
Fully documented training recipe
The card details data stages, bf16 mixed precision, AdamW schedules, and 64xH800 GPU-hours.
Use cases in the source record
- Step-by-step mathematical reasoning through chat-formatted prompts, as shown in the card's worked example.
- Short-form code generation such as utility functions, following the card's documented chat-format pattern.
Limitations and unknowns
- Benchmark figures are publisher-reported comparisons, not independently measured Ethen evaluations.
- Provider state is historical snapshot data, not independently refreshed current availability.
- The card notes the model inherits Llama-3.2 tokenizer and pre-training distribution and may reflect web biases.
Source and provenance
Source: InfiX-ai/InfiR-1B-Instruct
Captured: Unknown. Processed: 2026-09-07T19:35:06.588223+00:00.
Model Card for InfiR-1B-Instruct InfR aims to advance AI systems by improving reasoning, reducing adoption barriers, and addressing privacy concerns through smaller model sizes. Model Details Model Description Developed by: InfiX Language(s) (NLP): English Continual pretrained from model: [meta-llama/Llama-3.2-1B] Model Sources Repository: [github] Paper [optional]: [Arxiv] Uses Bias, Risks, and Limitations Performance gaps remain vs. 70 B+ models on very hard reasoning (e.g., OlympiadBench). Safety & bias : inherits Llama-3.2 tokenizer & pre-training distribution; may reflect web biases. Knowledge cut-off : mid-2023. Evaluation has…
F001F002F003F004F005F006F007F009F010F011F014F015F016F017F018F019F020