Real-time performance leaderboard
+Open-weight Moxie builds, merged and optimized with the Moxiegen + Method to run natively on consumer hardware. Pull one from Hugging Face and start chatting in + minutes — no data center required.
+8B params · GGUF · imatrix quant
+A three-way merge of Qwen 3.6, GLM + 5.2 and Gemma 4, optimized by the + Moxiegen Method to run natively on everyday consumer CPUs.
+$ llama serve -hf Moxiegen/Moxie-Cpu
+
+ Local OpenAI-compatible server with a web UI — or go
+ terminal-only with llama cli -hf
+ Moxiegen/Moxie-Cpu.
Install: curl -LsSf https://llama.app/install.sh | sh
+ (macOS/Linux) · winget install llama.cpp (Windows)
+
$ ollama run hf.co/Moxiegen/Moxie-Cpu
+
+ Also runs in LM Studio, Jan, vLLM & Docker.
+ Hugging + Face → +27B params · GGUF · BF16 weights
+A three-way merge of Qwen + 3.8, GLM 5.3 and Gemma 4, optimized by the Moxiegen Method to deliver + full-precision inference on consumer GPUs.
+$ llama serve -hf Moxiegen/Moxie-GPU:BF16
+
+ Local OpenAI-compatible server with a web UI — or go
+ terminal-only with llama cli -hf
+ Moxiegen/Moxie-GPU:BF16.
Install: curl -LsSf https://llama.app/install.sh | sh
+ (macOS/Linux) · winget install llama.cpp (Windows)
+
$ ollama run hf.co/Moxiegen/Moxie-GPU:BF16
+
+ Also runs in LM Studio, Jan, vLLM & Docker.
+ Hugging + Face →Both builds were created with the Moxiegen Method — lossless pointer-based weight mapping + that strips redundant weight so merged models run natively on the hardware you already own.