{"passport":{"slug":"keylm","display_name":"KeyLM","model_type":"language","parameter_count":"75M","context_window":null,"claimed_developer":"KeyLM","confirmed_developer":"KeyLM","developer_confirmed_at":"2026-06-04T23:36:46.749593+00:00","developer_source":"Confirmed via https://reddit.com/r/LocalLLaMA/comments/1tuyb8s/i_trained_a_75m_parameter_llm_from_scratch_on_18b/","first_appeared_at":"2026-06-04T23:36:46.749593+00:00","first_appeared_on":"reddit_localllama","status":"confirmed","availability":null,"deployment_status":"not_deployed","availability_scope":null,"gate":3,"eu_ai_act":"GPAI","regulatory_status":{"eu_ai_act":"GPAI"},"open_source":true,"weights_available":true,"license":null,"provenance_chain":[],"parent_model":null,"model_family":null,"superseded_by":null,"supersedes":[],"recommended_replacement":null,"product_links":null},"events":[{"event_type":"first_appearance","event_date":"2026-06-04T23:36:46.749593+00:00","title":"First recorded in AI news","detail":"KeyLM is a 75M parameter decoder-only language model trained from scratch on 18B tokens of public data. The instruction-tuned variant demonstrates competitive performance on instruction-following benchmarks (IFEval score of 17.85), outperforming SmolLM-135M-Instruct despite having less than half the parameters and significantly less training data. The model is available in pretrained base, instruction-tuned, and GGUF quantized versions.","source_url":"https://reddit.com/r/LocalLLaMA/comments/1tuyb8s/i_trained_a_75m_parameter_llm_from_scratch_on_18b/","source_name":"reddit_localllama"}],"delta":{"days_unattributed":0,"claimed_developer":"KeyLM","confirmed_developer":"KeyLM","match":true}}