Underdog-27B-1.0
Underdog-27B by Conway is a 27-billion-parameter model for conversation, writing, coding, and everyday assistance. Its MLX 4-bit QAT weights pair with Splash's speculative decoding for responsive local generation on Apple Silicon. We finetuned this model to run better in Underdog for human users, based on Qwen3.8.
Run with Splash
Install Splash 1.1.0 or later and authenticate with Hugging Face using an account with access to this repository. Then run:
hf auth login
splash serve --model ConwayResearch/Underdog-27B-1.0 --default-reasoning-effort medium
Splash automatically prepares the model and its compatible acceleration draft. Use the built-in chat interface or connect a compatible app. Medium reasoning is the recommended interactive default.
Local generation speed
| Task | MLX-LM | Splash + DFlash2 | Generation speedup |
|---|---|---|---|
| Email writing | 33.1 tokens/s | 80.6 tokens/s | 2.44x |
| Python coding | 33.0 tokens/s | 130.1 tokens/s | 3.94x |
| Long-document summarization | 32.5 tokens/s | 71.4 tokens/s | 2.20x |
Measured on an Apple M5 Max with 40 GPU cores and 48 GiB unified memory, macOS 26.5.1, using MLX 0.32.0 / MLX-LM 0.31.3 and Splash 1.1.0. Both configurations use the same MLX 4-bit weights. Values are mean decode rates from two runs per task, with temperature 0, reasoning set to none, a 512-token output budget, and no reused prompt tokens. Splash uses DFlash2 and INT8 KV; MLX-LM uses standard generation without a draft model. Engines were measured separately. Input lengths were 43, 44, and 3,003 tokens; generated output lengths may differ. Loading and warmup are excluded. Interactive medium reasoning is configured separately from these measurements.
Run with MLX
Download the complete repository and select its local directory in a compatible MLX runtime. Keep the supplied weights, tokenizer, configuration, and chat template together. The complete model download is approximately 16.08 GB; Splash manages its acceleration draft separately.
This release includes fine-tuning and quantization-aware training by Conway. The original license is preserved in LICENSE.