Skip to content
TheChatArchiveThe Chat Archive
o4-mini-2025-04-16· OpenAIHugging Face LMArena Human Preference 140K

which model is the best for multi turn? It should be open weight and less than 1B…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Metallama-4-maverick-03-26-experimental’s answer.
  1. User

    which model is the best for multi turn? It should be open weight and less than 1B parameters

  2. Assistant

    Here are a few fully open-weight, sub-1B models you can run locally—ranked by how well they handle multi-turn, chatty interactions:

    ModelSize (params)SpecializationProsCons
    DialoGPT-medium345 MMulti-turn dialogueTrained on Reddit conversations → decent continuity across turns. Very easy to run.Can sometimes drift off-topic.
    BlenderBot-small90 MSocial chatDesigned for open-domain chit-chat. Fast.Limited knowledge compared to larger.
    Flan-T5-large770 MInstructions + dialogueStrong at understanding instructions and maintaining context when prompted properly.Requires prompt-engineering to emulate “chatty” style.
    BLOOM-560m560 MGeneral-purpose LLMMultilingual. Works okay as a base for chat.Needs fine-tuning for best multi-turn flow.

    Recommendation:

    • If your main goal is chatty, back-and-forth conversation out of the box, go with DialoGPT-medium (345 M).
    • If you want a strong instruction-follower that you can steer into multi-turn chat via prompts, pick Flan-T5-large (770 M).

    Tips for improving multi-turn coherence

    1. Always prepend the full chat history (or a sliding window) to each prompt.
    2. Use a simple turn-tag format, e.g.:
      User: “…”
      Bot: “…”
    3. Experiment with sampling parameters (top-k≈40, top-p≈0.9, temperature≈0.7) to strike a balance between safety and diversity.

    Let me know if you want setup snippets or more on fine-tuning one of these for your specific domain! 😊

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.