philbert440/Qwen3.8-27B-W4A16-AWQ

Fix tokenizer: drop leftover truncation block, restore upstream tokenizer.json/tokenizer_config.json (fixes vision >1024 tokens; see discussion #3)

Measured performance v2: warm multi-pass methodology, draft-mode comparison, instruct + concurrency regimes

Update README.md

Model card v3: base-model benchmarks, official sampling params, YaRN long-context, citation

Model card v2: badges, V100 throughput chart, full validation matrix, serving recipes

Model card: recipe, thinking-mode calibration, V100 validation matrix w/ MTP

W4A16-AWQ g128 asym mse, June-proven hybrid smoothing, Magpie thinking calib, bf16 MTP graft

initial commit