Qwen3.8-Flash-Next · NInfer-Offload

English | 中文

Qwen/Qwen3.8-Flash-Next converted to the .ninfer v3 format for NInfer-Offload, which runs this 125B-total / 6B-active MoE model on a 16 GB consumer GPU by keeping the routed experts in system RAM. It includes the model's MTP predictor for speculative decoding.

This file only works with NInfer-Offload. It is not a GGUF or safetensors checkpoint and cannot be loaded by llama.cpp, vLLM, transformers or upstream NInfer.

Files

The 110 GB artifact is stored as NInfer's native multi-file set (at most 50 GB per file). Keep all files in one directory and pass only the entry file qwen3_8_flash_next.ninfer to NInfer; it reads the .part-000N continuation files itself. No reassembly is needed.

File Bytes SHA256
qwen3_8_flash_next.ninfer (entry) 49,999,998,976 9bb83f01…9bc529
qwen3_8_flash_next.ninfer.part-0001 49,999,998,976 8a115623…02ad05
qwen3_8_flash_next.ninfer.part-0002 9,593,264,896 a32e3c12…fff472
qwen3_8_flash_next.ninfer.conversion.json conversion report (format of every tensor)
SHA256SUMS full checksums

Requires NInfer-Offload v0.1.1 or newer; MTP needs v0.3.0 or newer. Older versions load the file and ignore the MTP predictor. v0.1.0 cannot map the PLE table across container files.

Updated 2026-10-02: the MTP predictor was added (1.5 GB); the text model is unchanged. The previous text-only files are replaced.

hf download luocha2050/Qwen3.8-Flash-Next-NInfer-Offload --local-dir flashnext
cd flashnext && sha256sum -c SHA256SUMS

The same files are mirrored on ModelScope: mymodel3861/Qwen3.8-Flash-Next-NInfer-Offload.

Run

Download the prebuilt Linux binaries from the GitHub release (RTX 40-series), or build from source. The simplest way downloads this model, plans GPU/CPU placement and opens a chat:

./ninfer-run Qwen/Qwen3.8-Flash-Next

Or run it directly (ninfer-serve takes the same options and serves an OpenAI/Anthropic API):

./ninfer qwen3_8_flash_next.ninfer --prompt "Hello" --no-thinking \
  --max-context 8192 --kv-capacity 8192 --spec mtp --draft-tokens 1 \
  --moe-offload --moe-threads 12 --moe-gpu-experts 64 --moe-expert-stats stats/stats_flash_next.txt

--moe-gpu-experts 64 keeps the 64 most frequently routed experts of each layer on the GPU (about 14.3 GiB VRAM); use 0–32 on GPUs with less free memory. --moe-threads is usually the number of physical CPU cores.

Requirements

  • NVIDIA RTX 40-series (sm_89, prebuilt binaries) or RTX 5090 (sm_120a, build from source). RTX 30-series and older are not supported.
  • x86-64 CPU with AVX2, FMA, F16C and BMI2.
  • About 80 GB of RAM (69 GB of routed experts are held in host memory); decode speed scales with memory bandwidth.
  • Linux. Only one machine has been tested (below).

Quantization

Tensors Format
Routed experts gate/up 4-bit, group 64, FP16 scales
Routed experts down 5-bit, group 64
PLE n-gram embedding table 5-bit
Token embedding 8-bit
Output head 6-bit
Other projections (attention, GatedDeltaNet, shared expert, hyper-connections) 8-bit, group 32
Router, shared-expert score, GatedDeltaNet a/b, norms BF16

Routed experts average 4.58 bits/weight (69.2 GB). The exact format of every tensor is listed in qwen3_8_flash_next.ninfer.conversion.json. The artifact was produced with python -m tools.convert --recipe qwen3_8_flash_next --components text,mtp from the official BF16 checkpoint; the MTP predictor uses the same formats.

Validation and speed

  • On a 59-token English paragraph, NInfer's mean NLL is 1.44330 (PPL 4.2346) versus 1.44474 (PPL 4.2407) from an independent FP32 reference over the same stored weights (tools/validate/qwen4exp_reference.py). In a teacher-forced greedy check, 15 of 16 tokens matched the reference argmax; the mismatch was a near-tie (0.29 nats).
  • No downstream benchmark or comparison with the BF16 model has been run.
  • Measured on a Xeon E5-2673 v3 (12 cores, ~21 GB/s memory read), 94 GB DDR3, RTX 4060 Ti 16 GB (PCIe 3.0 x8), greedy decode, BF16 KV, 4 prompts x 2 repetitions: 10.0–10.2 tok/s with 0 GPU experts per layer (5.2 GB VRAM), 12.5–14.1 tok/s with 64; prefill 22–35 tok/s.
  • With MTP and one draft per round (ninfer-serve, 64 GPU experts) the four prompts decode at 14.7–16.6 tok/s (3–22% faster than without MTP); two or three drafts help code (up to +39%) but slow down free-form prose. Draft acceptance ranged from 59% (essay) to 95% (code edit).

Limitations

  • Contexts above 2051 tokens use the model's QSA token selection and need --kv-dtype bf16 (the default) and text-only input.
  • Greedy output with MTP is not bit-identical to greedy output without it.

License

The model weights are a derivative of Qwen3.8-Flash-Next and are distributed under the Qwen Community License 1.0; the full text, including its copyright notice, is in LICENSE. Note its conditions on products above 100M monthly active users or US$20M monthly revenue, and its separate-license requirement for commercial Model-as-a-Service and AI Work Assistant businesses. NInfer-Offload itself is Apache-2.0.


中文说明

这是 Qwen/Qwen3.8-Flash-Next 转换成的 .ninfer v3 模型文件,供 NInfer-Offload 使用。 NInfer-Offload 把路由专家放在内存里,让这个总参数 125B、激活 6B 的 MoE 模型能在 16 GB 显卡上运行。 文件包含模型自带的 MTP 预测层,可用于投机解码。

本文件只能用 NInfer-Offload 运行,不是 GGUF 或 safetensors,llama.cpp、vLLM、transformers 和上游 NInfer 都无法加载。

下载

模型共 110 GB,以 NInfer 原生的多文件格式存放(每个文件不超过 50 GB):入口文件 qwen3_8_flash_next.ninfer 加续卷 .part-0001、.part-0002。所有文件放在同一目录,运行时只传入口文件, NInfer 会自动读取续卷,不需要合并。

需要 NInfer-Offload v0.1.1 或更新版本;使用 MTP 需要 v0.3.0 或更新版本。 旧版本也能加载本文件, 只是不使用 MTP。v0.1.0 无法跨文件映射 PLE 表。

2026-10-02 更新:加入 MTP 预测层(1.5 GB),文本模型部分不变,原来的纯文本文件已被替换。

hf download luocha2050/Qwen3.8-Flash-Next-NInfer-Offload --local-dir flashnext
cd flashnext && sha256sum -c SHA256SUMS

国内用户可从 ModelScope 下载同样的文件: mymodel3861/Qwen3.8-Flash-Next-NInfer-Offload。

运行

从 GitHub Release 下载预编译的 Linux 程序 (RTX 40 系),或自行编译。最简单的方式是一条命令:自动下载本模型、规划显卡/CPU 分工并进入对话:

./ninfer-run Qwen/Qwen3.8-Flash-Next

也可以直接运行(ninfer-serve 参数相同,提供 OpenAI/Anthropic 接口):

./ninfer qwen3_8_flash_next.ninfer --prompt "你好" --no-thinking \
  --max-context 8192 --kv-capacity 8192 --spec mtp --draft-tokens 1 \
  --moe-offload --moe-threads 12 --moe-gpu-experts 64 --moe-expert-stats stats/stats_flash_next.txt

--moe-gpu-experts 64 表示每层把最常用的 64 个专家放到显卡上(约需 14.3 GiB 显存),显存较少时用 0~32。--moe-threads 一般设为 CPU 物理核心数。

运行要求

  • NVIDIA RTX 40 系(sm_89,有预编译程序)或 RTX 5090(sm_120a,需自行编译)。不支持 RTX 30 系及更早。
  • 支持 AVX2、FMA、F16C、BMI2 的 x86-64 CPU。
  • 约 80 GB 内存(其中 69 GB 是路由专家);生成速度随内存带宽提高。
  • Linux。目前只在下面这一台机器上测试过。

量化格式

路由专家 gate/up 为 4 位,down 为 5 位(分组 64,FP16 缩放),平均 4.58 位/权重,共 69.2 GB; PLE n-gram 表 5 位;词嵌入 8 位;输出头 6 位;其他投影 8 位;路由器、归一化等为 BF16。 每个张量的格式见 qwen3_8_flash_next.ninfer.conversion.json。MTP 预测层使用同样的格式。

验证与速度

  • 59 个 token 的英文段落上,NInfer 平均 NLL 为 1.44330(PPL 4.2346),独立 FP32 参考实现为 1.44474(PPL 4.2407)。逐 token 贪心对比 16 个中 15 个一致,唯一不同处两个候选几乎并列(相差 0.29 nats)。
  • 尚未做下游评测,也没有和 BF16 原版对比质量。
  • 测试机:Xeon E5-2673 v3(12 核,内存实测读带宽约 21 GB/s)、94 GB DDR3、RTX 4060 Ti 16 GB (PCIe 3.0 x8)。贪心解码、BF16 KV、4 个提示各跑 2 次:每层 0 个 GPU 专家时 10.0–10.2 tok/s (显存 5.2 GB),64 个时 12.5–14.1 tok/s;预填充 22–35 tok/s。
  • 开启 MTP、每轮 1 个草稿(ninfer-serve,64 个 GPU 专家)时,4 个提示词解码 14.7–16.6 tok/s, 比不开 MTP 快 3%~22%;2~3 个草稿对代码更快(最多 +39%),写散文会变慢。草稿猜对率从 59%(作文) 到 95%(改代码)。

限制

  • 上下文超过 2051 个 token 时使用模型自带的 QSA token 选择,需要 --kv-dtype bf16(默认)且只支持纯文本输入。
  • 开启 MTP 后的贪心输出与不开时不完全一致。

许可证

模型权重是 Qwen3.8-Flash-Next 的衍生作品,按 Qwen Community License 1.0 分发,完整条款(含版权声明) 见 LICENSE。请注意其中对月活超过 1 亿或月收入超过 2000 万美元产品的要求,以及商业模型服务(MaaS) 和 AI 工作助手业务需另行取得授权的规定。NInfer-Offload 程序本身采用 Apache-2.0。

Downloads last month
53
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for luocha2050/Qwen3.8-Flash-Next-NInfer-Offload

Finetuned
(80)
this model