Qwopus3.8-27B-Coder-Flash-AWQ

In-house INT4 (AWQ-Marlin, W4A16, group size 128) build of Jackrong/Qwopus3.8-27B-Coder-Flash, a coding / agentic SFT of Qwen3.8-27B (Qwen3_5ForConditionalGeneration: 48 Gated-DeltaNet + 16 full-attention layers, 262K context, vision + video). Calibrated from the BF16 base with GPTQ on a thinking + image + video instruction mix (256 × 1024 tokens; AM-Thinking-v1, LLaVA-Instruct, LLaVA-Video-178K, NuminaMath-CoT, UltraChat), exported through compressed-tensors and converted to native AWQ. lm_head, the vision tower, the MTP layers and the DeltaNet gating scalars (in_proj_a / in_proj_b) stay BF16.

Recipe and serving stack: https://github.com/mattbucci/2x-3090-GA102-300-A1-sglang-inference (scripts/quantize/quantize_qwopus38_27b_coder_flash.py; serve with scripts/launch.sh qwopus38, SGLang v0.5.22 + the repo's patch series, 2× RTX 3090 at TP=2, full 262K context).

Gates on that stack (same instruments as the Qwen3.8-27B-AWQ ship): capabilities 5/5 (basic, tool call, thinking, image, video), needle 6/6 at 131K and 250K actual tokens, 256K reasoning probe 100 % at 131K and 256K, MMLU-57 50/57, HumanEval 30/30, scales audit clean (400 tensors, 0 flagged).