My First Model On @huggingface Qwen 3.8 27B Unleashed HIT 105,666 Downloads 🔥 Qwen3.8-27B Unleashed UD-Q3_K_XL hits near 100 tk/s @ 250k ctx on 1 x 4090. - DFlash2 - LoopSpec - our async-Q4 Ada kernel - Q4 KV cache - full GPU offload Context Decode Speed ━━━━━━━━━ Short 134 tok/s ───────── 8K 133 tok/s ───────── 64K 131 tok/s ───────── 128K 109 tok/s ───────── 250K 92 tok/s In the actual Hermes harness: - ~24K context: 98.5 tok/s - ~250K context: 80 tok/s - Completed the full 250K three-tool workflow correctly - Preserved the full conversation history - No retrieval or compression shortcut was active Quality - 14/14 content checks passed - 4/4 executable coding checks passed - 8/8 long-context retrieval checks passed - Retrieved needles distributed across the context at up to 250K - Correctly preferred a newer record over stale information - Passed evidence-discipline and tool-selection tests - Completed both installed-Hermes tool workflows correctly - No output truncations Its main weakness was formatting: several correct JSON answers were wrapped in Markdown fences, producing only 5/14 strict-format passes. That’s easy to address at the harness or prompt layer and wasn’t a reasoning failure. Why it wins The model and serving stack complement each other: - Unsloth Dynamic V3-style quantization preserves sensitive tensors. - Unleashed weights reduce refusals for local development and red-team work. - DFlash2 drafts target-specific token blocks. - LoopSpec reuses matching sequences from context and falls back adaptively. - The async-Q4 kernel improves Ada execution. - Q4 KV cache makes 262K context practical on 24GB VRAM. Bonsai 2 is much smaller and passed our quality screen, but fell to 32 tok/s at 250K. Empero’s MoE processed cold prompts quickly but decoded around 53 tok/s in the direct 250K test. Our winner sustained 92 tok/s in the matched direct test.