Repos made with Qwen3.8-Flash-Next

Qwen3.8-Flash-Next (W4A16) on 2x CMP 170HX: 110 tok/s decode, 7K prefill, 800 tok/s @ 23 concurrent - runs on 38 GB home RAM, no pinned memory

Qwen3.8-Flash-Next 180B NVFP4 @ 231 tok/s on one RTX PRO 6000 Blackwell (sm120) - sglang patches, launchers, full optimization analysis

LLM inference engine written from scratch in HIP for AMD Strix Halo (gfx1151) and Qwen3.8-Flash-Next - ~50 tokens/s out to 492k context

llama.cpp fork for GSQ-quantized Qwen3.8-Flash-Next on AMD Strix Halo (gfx1151): faster ROCm prefill, MTP and 4×256K chats with Q8_0 KV and one KV stream per slot, persistent KV cache on SSD. Fork of halo-box/strix-llama.cpp.

Qwen3.8-Flash-Next 177B MoE (ISTA GSQ-RCO IQ3_XXS) self-hosted on RTX 5090 Laptop (24GB VRAM + 64GB RAM) via Strata: up to 110 tok/s (100+ sustained), 256K full context, vision, MTP speculative decoding, reasoning default xhigh (max tier). GPU+CPU both saturated. 26 measured…

Run a 177B model (Qwen3.8-Flash-Next) on an 8 GB laptop GPU. Measured: 6.6 GiB VRAM, 47.8 GiB RAM, 34-35 tok/s. The n-gram table stays on disk; the experts run on CPU.

RTX 6000 96GB/Spark 128GB vLLM Recipe for Qwen 3.8 Flash Next

qwen3.8-flash-next on strix halo — 91g quant, 56 tok/s, 262k context, receipts included

Serving Qwen3.8-Flash-Next NVFP4 on a single DGX Spark (121 GiB) by paging its 51B n-gram embedding table to SSD swap

NVFP4 KV cache for Qwen3.8 Flash-Next QSA on SGLang (Pennyroyal)

Qwen3.8-Flash-Next on one DGX Spark: tuned vLLM launch, context sweep, tuning ladder. By @redp314

The best model under 200B at near 10 token/s exact batch throughput on a single laptop CPU, with an automatic 8 GB RAM path. Native C, no GPU or Python. | 200B 以下最强模型，单颗笔记本 CPU 精确批处理接近 10 token/s，最低 8 GB 内存自动运行。原生 C 语言，无需 GPU 或 Python。

Qwen3.8-Flash-Next optimized for 1x RTX Pro 6000

Agentic-coding oriented inference engine for N x V100-SXM2, based on SGLang: GLM-5.3-Flash, Qwen3.8-Flash-Next, DeepSeek-V4.1-Flash, MiniMax-H3.

CIRU llama.cpp runtime for Qwen3.8-Flash-CIRU-STRIX-IU4 on AMD Strix Halo

Run Qwen3.8-Flash-Next and Qwen3.8-27B on Apple Silicon Neural Engine

Getting Qwen's Qwen4-architecture preview (Qwen3.8-Flash-Next) to run on a single DGX Spark GB10 — 125.9 GiB of weights against 117 GiB of unified memory

Run Qwen3.8-Flash-Next (AtomicChat AD-4.27bpw Q4_K_M) with MTP on a llama.cpp box with 16 GB VRAM and 64 GB RAM: build, run flags, verification, measured numbers

Qwen3.8-Flash-Next on a Mac Studio M5 Ultra with oMLX: prompt-lookup MTP drafts (1.3-2.2x on edit turns, exact output), plus an engine comparison vs mlx-serve and TensorFold

fork of Niko1221/Strata: Qwen3.8-Flash-Next 177B(125B + Ngram 51B) on RTX 5060 Ti x4 + RTX 4070 SUPER x16 + i5-13500 + 48 GB. Tracking: xenodeve/Qwen3.8-Flash-Next-Tuning

Experimental: Qwen3.8-Flash-Next (180B/6B MoE) on one NVIDIA DGX Spark at 44 tok/s — 4-bit NVFP4 + FP8 hybrid under vLLM, with the quality cost measured and published

Verified Qwen3.8 Flash Next oMLX Lightning MTP deployment guide for Apple Silicon

Qwen3.8 Flash Next AWQ INT4 4x3090 v1: 262K, Vision, FP8 QSA KV and CUDA Graphs on 4x RTX 3090

This is an extension of perplexity.ai's lily inference engine, optimized to run Qwen3.8-Flash-Next on Apple Silicon (M5 or newer).