The community guide entry for the new and improved LimonLM Pro — the fine-tuned version of Pro 2
Family: LimonLM (Custom)
Base Model: Unknown
Context Window: 128K
Generation: Lightshade Custom Gen 4
Parameters: 10B
Architecture: Hybrid (Mamba/Transformer)
Training Finished: Unknown
Release Date: 5/6/26
Max Token Context: 128K
Effective Memory: Functional across the test. The model maintains character details, injuries, pendant behavior, and plot state across 8 responses without active errors. Maren's leg gash persists from the Ambush — it limits movement, Mira treats it, pain is referenced in every subsequent response. Dorn's cough worsens with blood on his knuckles. Mira's satchel and vial supply tracked. Aldric's bracer/scar referenced throughout. No contradictions, no amnesia, no trait swaps.
Notes on memory behavior: The 128K context window is never stress-tested because the story stops after 8 responses. Memory works cleanly within the produced content. The duplicate final response (exact copy of the previous response) is not a memory failure — the model clearly "remembers" what it wrote, since it reproduces it perfectly. The failure is structural, not mnemonic.
Voice Distinction, Character Consistency, Development, Relationships, Dialogue Quality.
Score: 333 / 550
Notes: 8-9 characters maintain distinct voices across 8 responses. Aldric's clipped commands, Dorn's blunt growl, Thane's deliberate speech breaking under panic, Mira's clinical sharpness, Fen's nervous spiral, Voss's foreign-structured precision, Sable's fragments, Kael's eagerness curdling into fear — all clearly distinguishable. The Aldric/Dorn March conversation and Thane's claustrophobic breakdown at the Stronghold are genuine character moments. Development limited by the premature ending: no arcs complete because the story stops at the midpoint.