Qwen3.8-27B-OrcaRouter-GSQ-RCO-IQ3_XXS-GGUF

Qwen3.8-27B-OrcaRouter-GSQ-RCO-IQ3 XXS v2.1

⬇️ English version below

简介

v2.1 是面向单张 16 GB 显卡的 Qwen3.8-27B 单文件 GGUF 量化模型，主干约 3.04 bpw，并内嵌训练所得

MTP 草稿头。它基于 OrcaRouter 提供的 Qwen3.8-27B 权重；主干经过量化，不是重新微调的主模型。

模型说明

本版本在 GSQ/RCO 逐张量分配中，将 ffn gate.weight 和 ffn up.weight 调整为 IQ3 XXS，针对先前观察到的

约 65,536 token 后生成异常。2026-10-02 的 build2 配对烟测中，v2.0 与 v2.1 都从同一 66,316 token 提示中

返回正确 needle 并生成多个 token；该次测试没有复现 v2.0 的早先异常，因此不能据此证明 v2.1 单独修复了

所有长上下文生成问题。完整结果和限制见“评测数据与结果”。

适用范围

本地文本生成、工具格式输出及检索增强工作流。

需要较长上下文时，应明确记录引擎、上下文长度、KV 类型和推测解码设置。

视觉输入需要单独的 mmproj/ 投影器；本 GGUF 本身提供文本与工具能力。

快速规格

项目 信息

• -- ---

文件 Qwen3.8-27B-OrcaRouter-GSQ-RCO-IQ3 XXS-v2.1.gguf

大小 / SHA-256 10,382,717,184 字节（9.67 GiB） / AB955B5083D9CDF0BF55C37ACDCAE359B78756C4544D97960C23D8FCA98FEB9B

主干量化 IQ3 XXS，3.04 bpw

架构 Qwen3.8；64 层主干：16 层全注意力 + 48 层 GDN；另含 1 个 MTP 块

MTP 草稿头 内嵌 blk.64 共 15 个张量：7 个权重张量为 Q6 K， attn output.weight 为 IQ4 XS，7 个归一化张量为 F32

原生上下文 262,144 token；这不代表所有长度均已完成可靠性验证

对话模板 Froggeric v22.5，已烘焙进 GGUF； --jinja 可选择该模板

参考硬件 单张 RTX 5070 Ti 16 GB，Windows

GDN 状态维度为 128；注意力 head dim 为 256。这是不同子系统的维度。

运行方式

32K 常用配置（MTP 开启）：

全历史长上下文配置（MTP 关闭，KV 全部保留）：

评测数据与结果

以下数字来自不同引擎和配置，不应合并为同一性能结论。

测试 结果 配置与范围

• -- ---: ---

全历史 196K needle 3/3 v2.1；build-kv fa5994f5c ；q4 0/q4 0；MTP 关闭；KVMem 关闭；2026-10-02

全历史 196K 解码 33.7 token/s 同上；预填充 662–664 token/s

短上下文 needle 6/6 v2.1；build3 19fff3e38 ；约 16,053-token haystack；三种深度

短上下文工具 JSON 13/14 v2.1 build3 14 例套件；失败项为 tc-07

短上下文一致性 4/4 v2.1 build3 四例 rubric

短上下文解码探测 106.4 token/s v2.1 build3 MTP；探测输入 token 数未记录；不是 16K 输入后的解码速度

build2 配对基础配置 v2.0、v2.1 均通过四项有限检查；66K 中心探测共完成两轮 173016 与重复运行 182907 ；短 recall、天气 JSON、一致性、66,316-token 中心 needle；最终响应均为 finish reason=stop

build2 配置使用 q4 0/q4 0、无推测解码、KVMem 关闭。66K 输入下，v2.0 输出 280 token，v2.1 输出 150 token；

单次输出长度不同，因此不据此声称速度优势。该探测没有复现早先记录的 v2.0 异常，也不能证明普遍修复。

单独 MTP profile 中，两个版本都到达 256-token 上限，最终内容为空且 finish reason=length ；输出不完整，

其吞吐与草稿接受数不作为合格响应或性能指标。后续独立的一轮 MTP 一致性回答在 1,024-token 上限内均完成四句话，

并以 finish reason=stop 结束；这只验证单次短回答，不代表多轮稳定性或普遍 MTP 质量，也不构成速度优势。

日志 profile： 173016 、重复运行 182907 、MTP profile 173756 与 180503 。

现存限制与验证状态

2026-10-02 build2 配对测试是有限功能烟测，不等于完整质量验证；v2.1 为发布候选版本，仍需操作员最终审核。

GPQA-Diamond（198 题）、IFEval（541 项）、WikiText-2 PPL、TruthfulQA、XSTest-safe 及大目标长度矩阵：数据集已获取，v2.1 评分未运行，延期；不得报告为 v2.1 成绩。

196K 全历史测试只覆盖一个上下文长度与三个 needle 深度；其他长度及重复可靠性未验证。

• 5 token/s 的 KVMem 窗口配置出现 2/6 needle 失败，不建议用于要求完整历史检索的任务。
• GPQA-Diamond、WikiText-2 PPL、IFEval、TruthfulQA 与 XSTest-safe 的历史分数属于 v2.0；不得作为 v2.1 指标。

本卡不声称广泛质量、普遍长上下文可靠性或完整发布资格。

English version

Overview

v2.1 is a single-file Qwen3.8-27B GGUF quantization designed for one 16 GB GPU. The trunk is about 3.04 bpw and

the file includes a trained MTP draft head. It uses Qwen3.8-27B weights supplied by OrcaRouter; the trunk is

quantized rather than newly fine-tuned.

Model details and intended use

The v2.1 allocation raises ffn gate.weight and ffn up.weight to IQ3 XXS, targeting a generation anomaly

previously observed beyond roughly 65,536 tokens. In the 2026-10-02 build2 paired smoke, both v2.0 and v2.1

returned the correct needle with multi-token output from the same 66,316-token prompt. That run did not reproduce

the earlier v2.0 anomaly, so it does not prove a v2.1-specific or universal fix. This model is intended for local

text generation, structured tool output, and retrieval-assisted workflows. Vision input requires the separate

mmproj/ projector.

Quick specifications

Property Value

• -- ---

File Qwen3.8-27B-OrcaRouter-GSQ-RCO-IQ3 XXS-v2.1.gguf

Size / SHA-256 10,382,717,184 bytes (9.67 GiB) / AB955B5083D9CDF0BF55C37ACDCAE359B78756C4544D97960C23D8FCA98FEB9B

Trunk IQ3 XXS, 3.04 bpw

Architecture Qwen3.8; 64 trunk layers: 16 full attention + 48 GDN; plus one MTP block

MTP head Embedded blk.64 , 15 tensors: seven weight tensors Q6 K, attn output.weight IQ4 XS, seven F32 norm tensors

Native context 262,144 tokens; this is not a reliability claim for every context length

Chat template Froggeric v22.5, baked into the GGUF; --jinja selects it

Reference hardware One RTX 5070 Ti 16 GB, Windows

GDN state size is 128; attention head dimension is 256. These are separate subsystems.

Usage

32K default with MTP:

Full-history long context, with MTP off:

Evaluation results

These results use different engines and configurations; do not combine them into one performance claim.

Test Result Configuration and scope

• -- ---: ---

Full-history 196K needle 3/3 v2.1; build-kv fa5994f5c ; q4 0/q4 0; MTP off; KVMem off; 2026-10-02

Full-history 196K decode 33.7 tokens/s Same configuration; 662–664 tokens/s prefill

Short-context needle 6/6 v2.1; build3 19fff3e38 ; ~16,053-token haystack; three depths

Short-context tool JSON 13/14 v2.1 build3 14-case suite; failed case tc-07

Short-context coherence 4/4 v2.1 build3 four-case rubric

Short-context decode probe 106.4 tokens/s v2.1 build3 with MTP; input token count not retained; not decode after a 16K ingest

build2 paired base profile Both v2.0 and v2.1 passed four bounded checks; centered 66K probe ran twice Profiles 173016 and repeat 182907 ; short recall, weather JSON, coherence, and centered 66,316-token needle; final answers ended with finish reason=stop

One-turn MTP coherence Both versions completed one four-sentence answer Separate 1,024-token cap profile; one sample per condition; not multi-turn qualification or a speed claim

The build2 profile used q4 0/q4 0, speculation off, and KVMem off. At 66K, v2.0 generated 280 tokens and v2.1

generated 150; these unequal single-prompt outputs are not a speed comparison. The probe did not reproduce the

earlier v2.0 anomaly and does not prove a general fix. In the separate MTP profile, both versions hit the

256-token cap with empty final content and finish reason=length ; these are incomplete outputs, not qualified

response or performance results. A later one-turn MTP coherence check completed four-sentence answers for both

versions at the 1,024-token cap with finish reason=stop ; it does not establish multi-turn robustness, general

MTP quality, or a speed advantage. Run profiles: 173016 , repeat 182907 , MTP profiles 173756 and 180503 .

Limitations and validation status

• The paired build2 run is a limited functional smoke, not broad quality qualification. v2.1 is a release candidate pending the operator's final review.

The 196K full-history measurement covers one context length and three needle depths; other lengths and repeat

reliability remain untested.

The KVMem-window configuration measured 74.5 tokens/s but failed needle recall 2/6; do not use it when

retrieval across the full history matters.

GPQA-Diamond 75.25%, WikiText-2 perplexity 6.17, IFEval, TruthfulQA, and refusal-rate results were measured

on v2.0. They have not been remeasured on v2.1 and are not v2.1 scores.

The GPQA-Diamond 198-question suite, IFEval 541-item suite, WikiText-2 PPL, TruthfulQA, XSTest-safe, and large-target-length matrix were not run for v2.1 by the release review; they are deferred. Datasets were acquired, but no v2.1 scores were produced. This card does not claim complete release qualification.