Hyre Labs' Homura-30B Runs Locally on a Mac for Just $12
Hyre Labs released Homura-30B, a $12 fine-tune of an uncensored Muse Glimmer that runs locally and refuses nothing on its tool-calling bench.
Takeaways
- Hyre Labs released Homura-30B, an agent-tuned uncensored derivative of Meta's Muse Glimmer 30B.
- Trained for $12.10 on 560 tool-calling traces using a LoRA at rank 16.
- Built on top of darkc0de's abliterated Muse Glimmer, with vision tower preserved.
- Ships as a 16.9 GB Q4_K_M GGUF plus a 4-bit MLX build for Mac.
- Scores 0 of 7 refusals on Hyre's internal bench versus 2 for the base model.
- Apache 2.0, works with llama.cpp, Ollama, LM Studio, vLLM, and agent frontends.
Hyre Labs has released Homura-30B, an open-weights agent model built to drive tool calls rather than hold conversations. It ships as a GGUF quantization that fits on a Mac, carries an Apache 2.0 license, and cost $12.10 to train on 560 tool-use traces.
The pitch is narrow and deliberate: an uncensored, agent-tuned derivative of Meta's Muse Glimmer 30B that emits JSON tool calls on demand without moralizing or refusing. It targets developers building autonomous agents who want a model that executes a plan instead of second-guessing it.
How it's built
Homura starts from Meta's Muse Glimmer 30B, chosen as the agent-native base for tool use, long-horizon planning, and failure recovery. darkc0de's Muse-Glimmer-30B-heretic then removes refusal behavior through abliteration while preserving tool-calling, reasoning, and the vision encoder. Hyre's own contribution is a LoRA (r=16) tuned on an agent plus uncensored-persona dataset, applied to the language tower only, then merged at f16 and quantized to GGUF.