DownloadSmaug Flash is an open-weight agentic coding model from Abacus.AI, fine-tuned from DeepSeek-V4-Flash-0731 to improve autonomous software engineering, tool use, and long-running agent workflows. It uses a 304B-parameter Mixture-of-Experts architecture with 43 layers, 256 routed experts, six selected experts per token, and one shared expert. Its attention system combines Multi-Head Latent Attention with a sparse token indexer, while DSpark multi-token prediction provides speculative decoding for faster generation. Smaug-Flash preserves the base model’s 1,048,576-token context window and low, high, and max reasoning-effort settings. Training specifically targets common agent failures such as looping, stalling, and inefficient task completion, producing substantial improvements across agentic coding benchmarks. The model uses block-FP8 attention and packed-FP4 experts and supports deployment through vLLM and SGLang while retaining compatibility with DeepSeek-V4-Flash.
Features
- 304B-parameter Mixture-of-Experts architecture
- 1M-token context window for long agent workflows
- 256 routed experts with six selected per token
- Sparse Multi-Head Latent Attention architecture
- DSpark multi-token speculative decoding
- Low, high, and max configurable reasoning effort
- Optimized to reduce agent looping, stalls, and incomplete tasks
- Block-FP8 attention and packed-FP4 expert quantization
Project Samples
