license: apache-2.0
library_name: transformers
base_model: xtools-at/Qwen3.8-27B-CodeForge-3xMerge
pipeline_tag: image-text-to-text
language:
- en
- zh
tags: - qwen3.8
- qwen3_8
- merge
- coding
- code-generation
- reasoning
- efficient-thinking
- agentic
- tool-use
- instruction-following
- distillation
- transformers
- safetensors
- heretic
- uncensored
- decensored
- abliterated
- reproducible
This is a decensored version of xtools-at/Qwen3.8-27B-CodeForge-3xMerge, made using Heretic v2.1.0.dev0
[!TIP]
This model is reproducible!See the README in the
reproducedirectory for more information.
Arbitrary-Rank Ablation (ARA) parameters
| Parameter | Value |
|---|---|
| start_layer_index | 32 |
| end_layer_index | 63 |
| preserve_good_behavior_weight | 0.9540 |
| steer_bad_behavior_weight | 0.0001 |
| overcorrect_relative_weight | 0.2984 |
| neighbor_count | 11 |
Model weights loaded in BNB 4bit quantization in Heretic. The resulting 16bit LoRA adaptor has been merged back into the BF16 model.
Performance
| Metric | This model | Original model (xtools-at/Qwen3.8-27B-CodeForge-3xMerge) |
|---|---|---|
| Refusals | 11/100 | 95/100 |
| KL divergence | 0.0128 | 0 (by definition) |
🛠️ Qwen3.8-27B Code Forge - 3x Merge
A three-way merge bringing together coding, reasoning, and distilled conversational capabilities.
Looking for quantizations/GGUFs? Click here
✨ Overview
This model is a weighted merge of three independently fine-tuned versions of Qwen3.8-27B. The idea behind the merge is fairly straightforward: combine different training approaches and strengths without leaning too heavily into just one specialization.
The three contributors bring complementary qualities:
- Advanced coding, structured reasoning, and efficient thinking from nerkyor's Coder390 EfficientThink.
- Conversational and knowledge-oriented distillation from TeichAI's Fable Distill.
- Coding and agentic capabilities from Jackrong's Qwopus Coder Flash.
The goal is a versatile general-purpose model that can write and understand code, work through challenging problems, follow instructions, and handle longer reasoning tasks while avoiding unnecessary reasoning overhead.
⚖️ Merge Details
The model was created by merging the following checkpoints with these weights:
| Model | Merge weight |
|---|---|
| nerkyor/Qwen3.8-27B-Coder390-EfficientThink-[...] | 1.00 |
| TeichAI/Qwen3.8-27B-Fable-Distill | 0.85 |
| Jackrong/Qwopus3.8-27B-Coder-Flash | 0.70 |
All three models share Qwen3.8-27B as their underlying foundation.
The weights above are the merge coefficients, not percentages of training data or a claim about each model's relative contribution to a particular capability.
🧩 The Three Contributors
1. 🧠 Coder390 EfficientThink - nerkyor
This is the most extensively post-trained contributor of the three. It builds on nerkyor's own EfficientThink SFT + SimPO checkpoint, itself derived from the official Qwen3.8-27B model.
Its training combines multiple rounds of supervised fine-tuning (SFT) and Reinforcement Learning from Offline/Online-style comparisons (RLOO, as described in the upstream recipe).
The training provenance is particularly interesting:
- Opus 5.5 and GPT-6 Astra: contributed gold answers and teacher trajectories.
- Grok 4.7: hosted the training run and helped check and filter data.
- DeepSeek V4 Pro and K3: contributed teacher trajectories.
- K3: additionally contributed problem writing and RLOO value review.
The EfficientThink objective is to reduce unproductive reasoning loops and runaway thought sequences without removing the ability to reason deeply when a problem actually requires it.
The Coder390 name reflects the author's stated GPQA, MMLU, and LiveCodeBench performance targets. The upstream model also includes MTP and DFlash2 speculative-decoding components.
This contributor brings additional reasoning refinement, coding specialization, and an explicit focus on efficient problem-solving.
2. 📖 Qwen3.8-27B-Fable-Distill - TeichAI
Fable Distill is a relatively light fine-tune of Qwen3.8-27B, trained on a combination of publicly identified Fable datasets and a larger private corpus of Fable 5 data.
The two publicly listed datasets are:
- armand0e/claude-fable-5-claude-code - Claude/Fable-oriented coding traces.
- armand0e/Fable-5-Chat - Fable 5 conversational examples.
The model's stated goal is to transfer useful Fable 5 conversational and problem-solving behavior into the Qwen3.8 architecture. Its published ARC Challenge and BoolQ results also suggest an emphasis on retaining and improving general knowledge and instruction-following capabilities.
This contributor adds a different flavor to the merge: conversational ability, knowledge-oriented responses, and distilled teacher behavior.
3. ⚡ Qwopus3.8-27B-Coder-Flash - Jackrong
A coding-focused fine-tune built on Qwen3.8-27B, with an emphasis on practical software engineering and agentic workflows.
Its training philosophy revolves around making each token count: understanding a task, planning an implementation, using tools, inspecting feedback, fixing errors, and delivering a working result.
The published training recipe describes approximately 200K code and reasoning examples, covering agent interactions, algorithmic problems, code generation, and logical reasoning. Its data sources include teacher-generated examples from Qwen3.8-Flash-Max, Qwen3.8-Max, DeepSeek-V4-Flash, and the NVIDIA Nemotron ecosystem.
This contributor is intended to bring strong implementation skills, code quality, tool-use behavior, and more purposeful reasoning.
🎯 Intended Use
The merged model is intended for:
- Software development, code generation, debugging, and refactoring.
- Complex reasoning, mathematics, and multi-step problem-solving.
- Agentic workflows involving tool use and iterative execution.
- General instruction following and conversational tasks.
- Technical explanations, analysis, and knowledge-intensive queries.
- Tasks where useful reasoning matters more than simply producing long reasoning traces.
It should be considered an experimental general-purpose merge rather than a model optimized for just one benchmark or task.
⚠️ Limitations
This is a weight-space merge, not a new round of supervised fine-tuning or reinforcement learning. Combining checkpoints does not guarantee that every capability of every contributor will be preserved.
Potential limitations include:
- Inconsistent behavior between different types of coding and reasoning tasks.
- Changes in response style or reasoning length.
- Reduced performance in some specialized tasks compared with an individual contributor.
- Tool-calling behavior that may depend on the inference engine and chat template.
- No independent benchmark results are claimed for this merged checkpoint.
As always, generated code and reasoning should be checked before being used in production or other consequential settings.
🙏 Credits and Acknowledgements
A big thank-you to the people behind the three upstream models and to the wider Qwen ecosystem.
Special thanks to:
- Qwen for the Qwen3.8-27B foundation model.
- Jackrong for the Qwopus coding and agentic fine-tune.
- TeichAI for Fable Distill.
- nerkyor for the EfficientThink and Coder390 training work.
- The contributors to the upstream datasets and teacher-model ecosystems.
This merge builds on their work rather than replacing it. Please visit and credit the original repositories when using or redistributing the model.
📜 License
The three listed upstream model repositories specify Apache-2.0. This model card uses the same license designation; users should review the upstream model and dataset terms and any applicable third-party terms before redistribution or commercial use.