1. Qwen 3.6 27B Abliterated Released to Reduce Refusal Rates
The Apostate project has launched its first large model release, an abliterated version of Qwen 3.6 27B. By modifying the model's safety alignment, the developers successfully reduced its refusal rate from 92% to 7.6%. The modification preserves the model's core capabilities, showing a minimal impact of only 0.120 KL divergence, and is available in both standard and GGUF formats on Hugging Face.
- • The Apostate project released Qwen 3.6 27B Abliterated on Hugging Face in standard and GGUF formats.
- • The model's safety alignment was modified to reduce the refusal rate from 92% to 7.6%.
- • The modification had a minimal impact on model capabilities, measured at 0.120 KL divergence.
- • This is the first large model release from the Apostate project.
Allows developers to deploy a highly capable local model that bypasses strict safety refusals while maintaining original capabilities.
2. Recall Plugin Adds Local Session Memory to Claude Code
A new plugin for Claude Code called Recall provides persistent session memory without sending data to external APIs. Operating entirely on the user's local machine, Recall uses a Python script implementing TF-IDF and TextRank to summarize project sessions. It writes logs and summaries to a local .recall/ directory, solving the cold-start problem for new coding sessions and reducing token usage. It also features built-in redaction to prevent API keys and secrets from being written to the logs.
- • Recall is a local-first plugin for Claude Code that creates session logs and condensed summaries of project sessions.
- • The tool runs entirely locally, using a Python script with TF-IDF and TextRank algorithms for summarization.
- • It saves API credits by reducing token consumption during session resumption and avoiding external model calls for summarization.
- • Recall includes a best-effort redaction feature to strip secrets like API keys and tokens before writing files.
- • The plugin supports both personal and shared team memory, storing data in a .recall/ directory.
Helps developers using Claude Code reduce API costs and avoid re-explaining project context by maintaining persistent, local session summaries.
3. Local Vision Model Benchmark Recommends Qwen3.6 27B
An updated local vision language model (VLM) benchmark evaluated 23 models using llama.cpp on Apple Silicon. The results show that Qwen3.6 27B (Q4, without thinking mode) leads the pack. Notably, the benchmark revealed that enabling thinking mode in hybrid models actually harms vision performance and causes timeouts, and that MoE models generally underperform compared to dense equivalents in vision tasks.
- • The benchmark evaluated 23 models across 30 images on an Apple M2 Max with 96GB RAM using llama.cpp.
- • Qwen3.6 27B (Q4, nothink) emerged as the top-performing model with a score of 79.6.
- • Enabling thinking mode in hybrid models was found to degrade vision performance, causing instability and timeouts.
- • Recommended models by VRAM tier include Qwen3.5 4B for 4-8GB, Qwen3-VL 8B (Q8) for 12-16GB, and Qwen3.6 27B for 24+GB.
- • Mixture of Experts (MoE) models performed worse than equivalent dense models on the vision tasks.
Gives developers clear, hardware-specific recommendations for deploying local vision models and warns of performance pitfalls like thinking-mode instability.