qwen36

Qwen3.6 27B × DFlash — 30-35 tok/s on NVIDIA DGX Spark (GB10) - LLama.Cpp

Minimal C++ LLM Qwen/Gemma inference engine

Run Qwen3.6 35B-A3B MoE locally on a 6 GB VRAM laptop

Experimental DDTree-on-vLLM research track for Qwen3.6 AEON Ultimate on DGX Spark / GB10.

Enables small-to-large self-hosted ai models to use local source code when running tool-calling agentic workloads. We actively data mine 20,900+ (2+ TB) popular github repos using large and small ai models to create reuseable: json, markdown and parquet files for local-first tool-calling models.

Run Qwen3.6-35B-A3B full native 258K context on 12GB VRAM (llama.cpp): ncmoe cliff rule, q4_0 KV prefill fix, 4 tuned profiles with scripts

Les 5 cercles + 18 stratégies d'optimisation tokens pour Qwen3.6+ (OpenRouter). Promptor v3 Council Edition : audit multi-perspective optionnel (5 advisors + peer review). Reverse Prompt Engineering. Python/PowerShell/OpenCode.

LLM inference benchmarks on Snapdragon X2 Elite Extreme (Adreno X2-90) with llama.cpp

Android port of Sonario for quickly summarizing YouTube videos, websites, documents, epubs & pasted text. Use qwen3.6-27b through Groq for fast free summaries, or run local LLMs for greater privacy.

Qwen3.6 27B MTP on Modal H100 with llama.cpp and an OpenAI-compatible API

Qwen3.6-27B at 111.6 tok/s (c4) on one DGX Spark—with FR-Spec, an x16 INT8 LM head, and 4.2× faster repeated-prefix TTFT.

Qwen3.6-27B on Ascend Atlas A2 with SGLang: BF16/W8A8 deployment notes, Mamba/NEXTN pitfalls, diagnostics, and reproducible examples.

Optimized Qwen3.6-35B-A3B recipe for DGX Spark — MTP Q4_0, 75-82 t/s

Run full 258K native context of Qwen3.6-35B-A3B MoE on 12GB VRAM with llama.cpp; benchmarked profiles, scripts, and tuning guide included.

Qwen3.6-35B-A3B on one NVIDIA DGX Spark (GB10, sm_121) with vLLM: FP8 vs NVFP4, MTP and DFlash speculative decoding, Marlin vs FlashInfer b12x MoE kernels, vLLM 0.23 vs 0.25. Measured tok/s, compose files, raw data. 日本語ノート / Notes in Japanese.