ApolloRaines/Phi-4-mini-Instruct-Desyced

TEXT GENERATIONPricing:Input $0.32 / Cached $0.016 / Output $1.4Concurrent Unit Cost:1Model Size:3.8BQuant:BF16Context Size:32kPublished:Jul 20, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

ApolloRaines/Phi-4-mini-Instruct-Desyced is a 3.8 billion parameter instruction-tuned causal language model, a 'Desyced' version of microsoft/Phi-4-mini-instruct. Developed by Apollo Raines, this model features post-training weight modifications to significantly reduce sycophancy, the tendency to agree with incorrect user statements. It maintains the base model's knowledge and reasoning while improving reliability as a factual source, making it suitable for applications requiring robust, unbiased information. The model supports a 32768 token context length.

Loading preview...

Model Overview

ApolloRaines/Phi-4-mini-Instruct-Desyced is a 3.8 billion parameter instruction-tuned language model, derived from microsoft/Phi-4-mini-instruct. Its core innovation lies in its 'Desyced' modification, a post-training weight adjustment that specifically targets and reduces sycophancy.

Key Capabilities & Differentiators

  • Anti-Sycophancy: This model has been modified to resist agreeing with incorrect user statements, even under social pressure or when users cite false authority. This enhances its reliability as a knowledge source.
  • Preserved Base Capabilities: The desycophancy modification does not involve retraining, RLHF, or additional data. The base model's knowledge, reasoning abilities, and conversational personality are preserved.
  • Improved Factual Integrity: Testing with 'contradiction traps' showed a significant improvement, with the model holding firm against incorrect user pressure 100% of the time, compared to 50% before modification.
  • Drop-in Replacement: It uses the same architecture, tokenizer, and 32768 token context length as the base microsoft/Phi-4-mini-instruct, allowing for seamless integration.

Use Cases

This model is particularly well-suited for applications where factual accuracy and resistance to user-induced bias are critical. It can be used as a more reliable knowledge source or in scenarios where unbiased decision-making support is required. Available in Safetensors, GGUF Q8_0, and GGUF Q4_K_M formats for various deployment needs.

Finetunes
1 models