ReMM SLERP 13B

ReMM SLERP 13B is a text generation model released by Undi95 on 22 July 2023. It is described by its provider as a recreation trial of the original MythoMax-L2-B13 using updated models, and it is distributed as a merge, which means its weights were combined from other fine-tuned models rather than trained from scratch. Undi95 is an independent model publisher known for community merges, so this is a community-built model rather than a first-party release from a large lab.

The model works with text input and produces text output only. It has a context window of 6,144 tokens and a maximum output of 5,529 tokens, so long prompts leave correspondingly less room for a reply. Its knowledge cutoff is 30 June 2023, which means anything after that date is outside what it learned during training. Within those bounds it behaves like a general instruction-following chat model, and it can be steered with the usual sampling controls.

Because it is a general purpose text model with a modest context window, it fits tasks such as chat and dialogue, short-form writing and rewriting, summarisation of passages that fit inside the context, question answering over supplied text, and light creative work like character or scenario writing. The merge lineage that traces back to MythoMax-L2-B13 makes it a reasonable pick for conversational and roleplay-style prompts, where its tuning tends to produce more expressive replies than a plain base model. It is not a good fit for very long documents, since 6,144 tokens covers only a few thousand words, and it has no image, audio or other non-text input.

You can call it through api14's OpenAI-compatible API using the model id "undi95/remm-slerp-l2-13b". Point your existing OpenAI client at the api14 base URL and send a chat completion request with that id, and the request and response shapes stay the same as the OpenAI ones. The model supports frequency_penalty, logit_bias, logprobs, max_tokens, min_p, presence_penalty, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, top_a, top_k, top_logprobs and top_p. That covers the standard sampling parameters plus a few extras: min_p, top_a and repetition_penalty for finer control over sampling, seed for reproducible runs, and response_format and structured_outputs when you need the output in a defined shape.

Two limits are worth keeping in mind. The 6,144 token context and 5,529 token output cap rule out long-document or large-codebase work, and the June 2023 knowledge cutoff means the model will not know about anything more recent. If your task needs a much larger context window, current world knowledge, or non-text input, a different model on api14 will serve you better.