Lernex/Metis-1.4-think

Metis-1.4 Think

Metis-1.4 Think is the reasoning-style SFT release from Lernex's Metis-1.4 research run: a compact ~500M-parameter MoR-style language model and a practical milestone toward the broader Metis efficient-model stack.

This checkpoint starts from Metis-1.4 Chat and adds the Think SFT stage, teaching the model to produce more reasoning-shaped completions. Reward modeling and DPO were intentionally skipped for this release so the finished Chat and Think SFT checkpoints could be preserved, published, benchmarked, and studied without spending more compute on low-expected-return polish.

Metis-1.4 Think is still a small model. It can emit legible reasoning-like traces, but it is not expected to be strong on hard math, code, or expert knowledge. Treat it as a research artifact and benchmark target, not a dependable assistant.

Files

• model.safetensors
• config.json
• generation_config.json
• tokenizer.json
• tokenizer_config.json
• special_tokens_map.json

Intended Use

Use this model for reasoning-style prompt probes, small-model evals, visible trace behavior checks, and comparisons against the Chat SFT release. The raw model output may include visible reasoning-like text depending on the prompt and decoding settings.

Release Note

This is the corrected Metis-1.4 Think SFT release: the model that never quit, and a pivotal step in Lernex's research toward more efficient learning-focused language models.