Agents-A1-Flash (Tencent STQ1_0 + FP8 Mixed)

Compressed release of InternScience/Agents-A1 (35B MoE, 256 experts, 8 active per token).

Compression Specification

• Routed MoE Experts: Tencent STQ1_0 (Sherry 3:4 fine-grained ternary sparsity, 1.3125 bpw).
• Attention & Shared Experts: FP8 (float8_e4m3fn).
• Layer Norms & Vectors: FP16.
• Original Size: ~70 GB BF16
• Compressed Size: 8.19 GB (88.3% reduction)
• Serving Budget: Runs in <3.5 GB VRAM using on-demand storage caching or ~8.2 GB full VRAM on an NVIDIA T4 GPU.

Shard Distribution

• model-00001 to model-00011 (~441 MB each): STQ1_0 routed MoE expert layers.
• model-00000, model-00012, model-00013: FP8 embeddings, vision/dense layers, and attention blocks.