Qwen3.8-Flash-Next-FLE-GGUF

Model overview

Three hybrid GGUF quantizations of Qwen3.8-Flash-Next (512 routed experts, 10 active per token) plus the model's own multi-token-prediction (MTP) head with its experts requantized to IQ4_NL for FLE. We also tried re-training the head; it did not pay off on unseen prompts, so the untrained head is what we ship (see MTP head). A pruned variant of the first, FLE-IQ3_M-REAP448, is in its own repo: sirfyyn/Qwen3.8-Flash-Next-FLE-REAP448-GGUF. The names follow the expert bit width, as in Unsloth's "UD" mixtures: the `IQ3_M`, `IQ3_XXS` and `IQ2_M` part says how many bits the routed experts get (see Bit allocation), and the `FLE-` prefix marks the mixture. These are not standard llama.cpp file types (`ftype`); every file mixes several tensor types. They were built and measured for **FLE (Fast Local Experts)**, our llama.cpp fork for running large MoE models on one GPU plus system RAM: . The three quantizations here and FLE-IQ3_M-REAP448 are standard GGUF files and load in upstream llama.cpp (FLE-IQ3_M-REAP448 simply has 448 instead of 512 experts per layer; its KLD below was measured with upstream). The speed numbers below, and the MTP head, need the FLE engine (see Running).

Key features

特点摘要待依据来源补齐，不凭模型名称推断能力。

Limitations & considerations

Sources are insufficient to establish specific limitations. Verify task performance, latency, data policies, licensing and version lifecycle.

Source prices & cost analysis

Prices vary by channel, region and configuration. Channel quotes are listed independently and are not treated as a single official provider price.

Public benchmarks & run configuration

暂无已采集的可核验公开评分。模型未上榜不代表能力较低。

Frequently asked questions

Sources & traceability

许可证：other。模型介绍与第三方评测均以原来源的具体版本和方法为准。