Jackrong/Qwopus3.8-27B-Flash-GGUF — new model trending #26 on Hugging Face
Community fine-tune Qwopus3.8-27B-Flash, built on Qwen3.8-27B, cuts agent reasoning latency with 12.8% faster decoding and 80.7% MTP acceptance.
Jackrong released Qwopus3.8-27B-Flash, a fine-tune of Qwen3.8-27B optimized for long-running agent workloads, reporting 12.8% faster decoding and 80.7% multi-token-prediction acceptance. Training used roughly 1.5 million teacher-scored SFT examples filtered to the top 10%, followed by reinforcement training with NVIDIA NeMo-RL and GSPO. The author notes an explicit trade-off: MMLU-Pro mixed-set scores are lower than the base model, and a known bug can produce incorrect Python indentation. Author-provided benchmarks have not been independently verified.
- Fine-tuned from Qwen3.8-27B for cheaper, faster agentic reasoning
- Two-stage pipeline: filtered SFT then NeMo-RL + GSPO reasoning RL
- Eval ensemble included Qwen3.7-Max, GLM-5, GPT-OSS-120B-High, Gemma4-27B
- Known issue: occasional wrong indentation in Python coding tasks
- MMLU-Pro mixed-set score lower than base model
Full article2,958 words · extracted from huggingface.co · click to collapse
# 🪐 Qwopus3.8-27B-Flash
<div align="center">
<img src="https://cdn-uploads.huggingface.co/production/uploads/66309bd090589b7c65950665/GvN4SANiFepBuWcsoWbOX.jpeg" alt="IMG_7359" width="82%"/>
</div>
> [!NOTE]
> **Qwopus3.8-27B-Flash** is a fine-tuned model built on [Qwen3.8-27B](https://huggingface.co/Qwen/Qwen3.8-27B). It is designed to retain strong general capability while substantially lowering reasoning cost and response time for practical, long-running agent workloads.
> [!TIP]
> **12.8% Faster Decoding · 80.7% MTP Acceptance · Less Runaway Reasoning · More Efficient Agent Completion**
> [!WARNING]
> **Known Issue**
>
> In some specific Python coding tasks, the model may occasionally generate code with incorrect indentation. A fix is currently in progress.
>
> **This issue is limited to Python code indentation in certain cases and does not affect the model's other capabilities or overall performance.**
## 💡 1. Fine-Tuning Philosophy — Preserve Capability, Collapse Cost
Qwopus3.8-27B-Flash starts from a simple premise: a useful Flash model should preserve enough capability to complete demanding work while driving down the cost and latency of reasoning.
An agent amplifies reasoning cost. A normal chat may call a model once, but one agent task can call it dozens or hundreds of times through a repeated loop:
```text
Read → Think → Tool Call → Observe → Edit → Test
```
> [!IMPORTANT]
> **Wall-clock time is what users feel.** If a model adds five seconds to each turn, a 50-turn task adds 250 seconds of waiting. In the agent era, every-token speed, reasoning length, and completion efficiency are as important as a single-turn benchmark result.
Inference cost is also a central commercialization constraint. Every generated token consumes GPU time, electricity, memory capacity, and concurrency budget. Flash models are intended to give users with constrained resources a more practical experience, not merely to produce a higher throughput number on paper.
The goal of Qwopus3.8-27B-Flash is therefore to reduce ineffective computation and reach a clean completion more quickly and consistently. The benchmark results below show that this optimization has an explicit trade-off: it improves inference efficiency, while the reported MMLU-Pro mixed-set score is lower than the base comparison.
### Five-Story Pagoda Garden — Visual Output Comparison
> [!NOTE]
> These visuals are qualitative model outputs. They complement the measured benchmark results below and are not a quantitative capability score on their own.
#### Original Base Model — Qwen3.8-27B
#### Fine-Tuned Model — Qwopus3.8-27B-Flash
## 🧪 2. Fine-Tuning Cookbook
### 2.1 Base Model and Training Goal
Qwopus3.8-27B-Flash is fine-tuned from the Qwen3.8-27B foundation model. The training objective is not simply to maximize visible reasoning length or claim a universal benchmark gain. It is to retain practical problem-solving ability while making the model faster, less prone to pathological long-tail reasoning, and better suited to iterative agent workflows.
### 2.2 Stage 1 — Data Preparation and Quality SFT
Data preparation is the foundation of the SFT stage. The first stage began with approximately **1.5 million** teacher-model SFT examples. After extensive cleaning and filtering, the highest-quality **10%** was retained for the initial fine-tuning stage.
Each example was evaluated across its three main components—**question**, **chain of thought**, and **answer**—with criteria tailored to a 27B model:
- Semantic relevance and usefulness.
- Problem difficulty.
- Chain-of-thought quality.
- Answer consistency.
The evaluation ensemble used reasoning models including **Qwen3.7-Max**, **GLM-5**, **GPT-OSS-120B-High**, and **Gemma4-27B**. Their scores were combined through a weighted calculation, after which only high-quality examples were retained.
> [!NOTE]
> The training mix also includes agent-trajectory data and reconstructed trace data derived from closed models such as Claude and GPT. Detailed examples cannot currently be disclosed. The dataset will be further organized and released after preparation is complete.
### 2.3 Stage 2 — NeMo-RL + GSPO Reasoning Reinforcement
Stage 2 consolidates and strengthens the chain-of-thought and reasoning behaviors learned during SFT. This stage follows an **NVIDIA NeMo-RL + GSPO** route for reasoning reinforcement.
The training process uses repeated sampling and reward comparison, then updates the model with a sequence-level importance ratio. The intent is to reinforce useful reasoning trajectories and completion behavior rather than reward longer reasoning traces by default.
> [!TIP]
> The two stages serve different roles: Stage 1 builds a high-quality reasoning and instruction-following foundation; Stage 2 consolidates it through reward-guided reasoning reinforcement.
## 🙏 3. Training Stack, Testing, and Collaboration
### Unsloth
Special thanks to [Unsloth](https://unsloth.ai/) for its efficient, memory-optimized fine-tuning framework. Its tooling made the large-model training workflow more practical and accessible.
### Kyle Hessling
Special thanks to [**Kyle Hessling**](https://x.com/KyleHessling1) for completing the testing work and providing essential evaluation support for this release.
> [!TIP]
> If you have any questions or suggestions about the model, please feel free to reach out to [Kyle Hessling](https://x.com/KyleHessling1) on [X](https://x.com/KyleHessling1) and share your feedback. Thank you so much for all your support! 🙏
> [!NOTE]
> The training and test results in this card are author-provided local results. The sections below describe the measured scope and do not claim universal behavior across all hardware, prompts, or agent environments.
> [!WARNING]
> ⚠️ **Known Issue**
>
> In some specific Python coding tasks, the model may generate code with incorrect indentation. A fix is currently in progress.
>
> **This issue is isolated to Python code indentation and does not affect the model's other capabilities.**
## 📊 4. Benchmark Results
| Reported signal | Qwopus3.8-27B-Flash | Qwen3.8 base comparison | Interpretation |
| --- | ---: | ---: | --- |
| MMLU-Pro mixed-question accuracy (2,500 questions) | 91.28% | **92.73%** | **−1.45 pp** |
| Decoder-only throughput across Math, Physics, and Chemistry | **9.347 tok/s** | 8.284 tok/s | **+12.8%** |
| Weighted MTP draft acceptance | **80.7%** | 66.1% | **+14.6 pp** |
| Aggregate `raw_output` characters across three subjects | **7,949,546** | 8,824,213 | **−9.9%** |
| End-to-end batch evaluation throughput | **~8% higher** | — | Separate, approximate measurement |
| Agentic software-engineering battery, strict one-run result | **13 / 14 (93%)** | — | 26.0 min on one RTX 5090 |
### 4.1 Evaluation Scope
The reported local tests used inference environments that included NVIDIA GeForce RTX 5090, NVIDIA V100, and NVIDIA GB10 hardware. The current test brief does not map every table or case study to a particular device.
For the MMLU-Pro decoder-only comparison, both models used the same Q5_K_M + MTP configuration:
```text
draft_n_max = 2
-c 327680
-np 10
```
The agentic software-engineering battery in Section 4.6 is a separate run explicitly reported on one RTX 5090 and uses its own configuration, listed with that battery.
### 4.2 MMLU-Pro Accuracy Reference and Five-Layer Tower Renders
> [!WARNING]
> **Accuracy–efficiency trade-off:** The fine-tuned model improves reasoning efficiency, but its reported score on the 2,500-question MMLU-Pro mixed set is **91.28%**, compared with **92.73%** for the Qwen3.8 base comparison—a decrease of **1.45 percentage points**. The decoder-only throughput results below measure speed, not answer accuracy.
### 4.3 MMLU-Pro Decoder-Only Throughput
This table measures isolated generated-token throughput rather than answer accuracy. Evaluation time is in seconds; throughput is generated tokens per second.
| Subject | Qwopus generated | Qwopus eval (s) | Qwopus tok/s | Qwen generated | Qwen eval (s) | Qwen tok/s | Gain |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| Math | 753,406 | 81,212.1 | **9.277** | 804,071 | 96,614.0 | 8.323 | **+11.5%** |
| Physics | 1,055,509 | 111,465.3 | **9.469** | 1,021,881 | 125,039.7 | 8.173 | **+15.9%** |
| Chemistry | 1,428,542 | 153,685.2 | **9.295** | 1,372,810 | 164,461.4 | 8.347 | **+11.4%** |
| **Three-subject total** | **3,237,457** | **346,362.6** | **9.347** | **3,198,762** | **386,115.2** | **8.284** | **+12.8%** |
Across the three reported subjects, Qwopus reaches **9.35 tok/s** versus **8.28 tok/s** when rounded to two decimals. Every reported subject shows a double-digit decoder-only throughput gain. The separately reported ~8% end-to-end batch-throughput result includes the full evaluation path and should not be combined with this decoder-only measurement.
### 4.4 MTP Draft Acceptance
Draft acceptance is an internal llama.cpp/MTP implementation metric. The reported aggregate is a weighted calculation, rather than an average of subject-level percentages:
```text
sum(accepted) / sum(generated)
```
| Subject | Qwopus3.8-27B-Flash | Qwen3.8 base comparison | Difference |
| --- | ---: | ---: | ---: |
| Math | **81.1%** | 66.3% | **+14.8 pp** |
| Physics | **79.2%** | 64.6% | **+14.5 pp** |
| Chemistry | **81.7%** | 67.1% | **+14.6 pp** |
| **Three-subject weighted total** | **80.7%** | **66.1%** | **+14.6 pp** |
| Three-subject raw counts | Accepted | Generated |
| --- | ---: | ---: |
| Qwopus3.8-27B-Flash | 1,998,599 | 2,476,486 |
| Qwen3.8 base comparison | 1,821,083 | 2,755,425 |
> [!TIP]
> The **+14.6 pp** result is an absolute difference in percentage points. It is not a +14.6% relative-improvement claim.
### 4.5 Reasoning-Efficiency Findings
The following values are `raw_output` **character counts**, not token counts. Mean, P95, and median describe different parts of the output-length distribution.
| Subject | Mean Qwen | Mean Qwopus | Mean change |
| --- | ---: | ---: | ---: |
| Math | 4,614.2 | 3,798.7 | **−17.7%** |
| Physics | 5,995.0 | 5,469.6 | **−8.8%** |
| Chemistry | 7,039.2 | 6,630.8 | **−5.8%** |
| Aggregate `raw_output` characters | Qwen3.8 base comparison | Qwopus3.8-27B-Flash | Change |
| --- | ---: | ---: | ---: |
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/Jackrong/Qwopus3.8-27B-Flash-GGUF