OliviaRossi/Qwopus-KAT-Coder-35B-Merged-GGUF overview
license: apache 2.0 language: en zh pipeline tag: text generation tags: code agent agentic coding moe gguf llama cpp apex quality q4 k m q8 0 reasoning tool ca…
Runs locally from ~1.66 GB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
Model Details
| Model ID | OliviaRossi/Qwopus-KAT-Coder-35B-Merged-GGUF |
|---|---|
| Author | OliviaRossi |
| Pipeline | text-generation |
| License | apache-2.0 |
| Base model | OliviaRossi/Qwopus-KAT-Coder-35B-Merged,Jackrong/Qwopus3.6-35B-A3B-Coder,Kwaipilot/KAT-Coder-V2.5-Dev |
| Last modified | 2026-08-27T13:52:19.000Z |
Model README
---
license: apache-2.0
language:
- en
- zh
pipeline_tag: text-generation
tags:
- code
- agent
- agentic-coding
- moe
- gguf
- llama-cpp
- apex-quality
- q4_k_m
- q8_0
- reasoning
- tool-calling
- software-engineering
- mtp
- multi-token-prediction
- vision
- multimodal
- image-to-code
base_model:
- OliviaRossi/Qwopus-KAT-Coder-35B-Merged
- Jackrong/Qwopus3.6-35B-A3B-Coder
- Kwaipilot/KAT-Coder-V2.5-Dev
---
<a id="readme-top"></a>
<div align="center">





Self-speculative Multi-Token Prediction, native vision input, and MoE Adaptive Precision (APEX) quantization — see capabilities and quantization details below.
</div>
<br>
Contents — At a Glance · Executive Summary · MTP & Vision · Fusion Architecture · Intended Use & Limitations · Quantization & Hardware Sizing · Quickstart · Sampling · Prompt & Chat Templates · Lineage · Citation · License
---
At a Glance
| | |
| :--- | :--- |
| Total parameters | ~35B |
| Active parameters / token | ~3B (8 of 256 routed experts + 1 shared expert) |
| Layers | 40 |
| Attention | Hybrid GatedDeltaNet linear attention with periodic full attention |
| Context window | 32,768+ tokens |
| Flagship Quantization | APEX Quality (~21.3 GB) · Q4_K_M (~20.4 GB) · Q8_0 (~35.2 GB) |
| Recommended Template | Improved Chat Template for Qwen 3.x |
| Ideal hardware | Single 24 GB GPU (RTX 3090/4090), or 32 GB+ unified memory (Apple Silicon) |
| Languages | English, Chinese |
| License | Apache 2.0 |
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
Executive Summary
Qwopus-KAT-Coder-35B is a Mixture-of-Experts (MoE) coding model built by merging two top-tier 35B fine-tunes:
- Jackrong/Qwopus3.6-35B-A3B-Coder — deep algorithmic logic synthesis, complex mathematical reasoning, and MTP speculative decoding.
- Kwaipilot/KAT-Coder-V2.5-Dev — state-of-the-art SWE-agent repository execution, native tool-calling, and automated bug resolution.
The result pairs multi-step reasoning depth with autonomous environment tool execution: a private, self-hosted coding agent that fits on a single 24 GB consumer GPU.
> Quantization Highlight: This repository features APEX Quality quantization — an adaptive mixed-precision MoE layout that preserves critical shared experts and attention blocks at high precision while applying layer-gradient compression across the 256 routed experts, delivering near-Q8 accuracy at a 21.3 GB footprint.
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
MTP & Vision Capabilities
Both parent lines trace back to Qwen3.6-35B-A3B (Alibaba Cloud), inheriting access to native vision encoding and a multi-token prediction (MTP) speculative head.
Multi-token prediction (MTP)
Self-speculative decoding allows the model to draft multiple tokens per forward pass using its own MTP head without requiring an external draft model. Enable it in llama.cpp via --spec-type draft-mtp.
- Realistic speedup: Community testing on Qwen3.6-35B-A3B MTP builds reports 1.4–2.2× faster generation with no accuracy loss, depending on prompt predictability and acceptance rates.
- Merge behavior note: The Qwopus3.6 parent ships an MTP head, while
KAT-Coder-V2.5-Devdoes not. The merged weights preserve the MTP projection layers; benchmark--spec-type draft-mtpagainst standard autoregressive decoding on your specific workload.
Vision / multimodal input
The underlying architecture supports image input for tasks like UI-screenshot-to-code, architecture diagram scaffolding, and GUI bug root-cause analysis.
> Requirement: Vision processing requires a matching multimodal projector (mmproj) file passed via --mmproj <path> (e.g. llama-mtmd-cli or llama-server). The text GGUF alone processes text and code tokens.
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
The Fusion Architecture
<div align="center">
</div>
| Highlight | Description |
| :--- | :--- |
| Active-parameter efficiency (A3B) | 35B total parameters with ~3B active per token (8 of 256 routed experts + 1 shared expert). Achieves 70B-tier coding performance at 3B-tier inference latency. |
| Hybrid GatedDeltaNet linear attention | Linear recurrence layers interleaved with periodic standard self-attention, keeping KV cache growth and memory scaling flat over long context (32k+ tokens). |
| Dual-format expert consolidation | Unified blending across Qwopus's fused 3D expert tensors and KAT-Coder's discrete expert matrices, combining logic reasoning and tool use into single cohesive expert representations. |
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
Intended Use & Limitations
Intended use
- Local, self-hosted coding agent: multi-file repository refactoring, bug fixing, test generation, and autonomous terminal tool execution.
- High-performance private code assistant running entirely on local consumer hardware (24 GB VRAM or 32 GB+ Apple Silicon).
- Multimodal UI-to-code workflows (when paired with a compatible
mmprojvision file).
Out of scope
- Unreviewed execution in safety-critical production environments without automated validation/sandboxing.
- General-purpose casual chat (the model is strongly optimized for coding, reasoning, and agentic JSON/XML tool calling).
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
Quantization & Hardware Sizing
Available Quantizations
| Quant Type | File Size | Memory Required (VRAM / RAM) | Precision Strategy | Recommended Setup |
| :--- | :---: | :---: | :--- | :--- |
| APEX-Quality | ~21.3 GB | 24 GB VRAM / 32 GB RAM | Adaptive MoE Gradient (Q8/Q6/Q5/Q4) | ⭐ Recommended Flagship (RTX 3090/4090, Mac 32GB+) |
| Q4_K_M | ~20.4 GB | 24 GB VRAM / 32 GB RAM | Standard uniform K-quantization | 24 GB VRAM / Apple Silicon |
| Q8_0 | ~35.2 GB | 40 GB+ VRAM / 64 GB RAM | Full 8-bit uniform quantization | A100 / A6000 / Mac Studio (64GB+) |
Deep Dive: APEX Quality MoE Architecture
Standard uniform quantization treats all weights equally, which can degrade sparse MoE networks. APEX (Adaptive Precision for EXpert Models) applies mixed-precision quantization mapped to the structural importance of each layer group:
- Shared Experts (
shexp) $\rightarrow$Q8_0: Always active across every token; kept at 8-bit precision to eliminate cumulative outlier degradation. - *Attention & SSM Projections (
attn_qkv,ssm_) $\rightarrow$Q6_K**: Preserves state-space recurrence and linear attention dynamics over long repositories. - Embeddings & Output (
token_embd,output) $\rightarrow$Q8_0/Q6_K: Maintains vocabulary coordinate stability and clean token sampling. - *256 Routed Experts (
ffn__exps) $\rightarrow$ 3-Tier Layer Gradient**:
- Edge Layers (L0–L4 & L35–L39): Q6_K (preserves initial feature extraction and final token routing).
- Near-Edge Layers (L5–L9 & L30–L34): Q5_K (smooth transition gradient).
- Middle Layers (L10–L29): Q4_K / IQ4_XS (aggressive compression across redundant sparse experts).
> Result: APEX Quality matches uncompressed Q8_0 coding benchmarks while fitting entirely within 24 GB VRAM at 30–60+ tokens/sec.
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
Quickstart Guide
1. Launch with llama.cpp
Interactive CLI (APEX Quality + MTP speculative decoding)
llama-cli \
-hf OliviaRossi/Qwopus-KAT-Coder-35B-Merged-GGUF:Qwopus-KAT-Coder-35B-Merged-APEX-Quality.gguf \
-ngl 99 \
-c 32768 \
--spec-type draft-mtp \
--temp 0.7 --top-p 0.95 --top-k 20 \
-p "You are an expert autonomous software engineer. Write a production-grade Async Task Queue in Python with Redis backend."
Local OpenAI-Compatible Server
llama-server \
-hf OliviaRossi/Qwopus-KAT-Coder-35B-Merged-GGUF:Qwopus-KAT-Coder-35B-Merged-APEX-Quality.gguf \
-ngl 99 \
-c 32768 \
--spec-type draft-mtp \
--temp 0.7 --top-p 0.95 --top-k 20 \
--port 8080 \
--host 0.0.0.0
> Vision Support: To enable multimodal image input, append --mmproj <path-to-mmproj.gguf> to your llama-server command.
<details>
<summary><b>2. Ollama Integration (with Custom Modelfile)</b></summary>
<br>
Create a Modelfile:
FROM ./Qwopus-KAT-Coder-35B-Merged-APEX-Quality.gguf
TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{- range .Messages }}<|im_start|>{{ .Role }}
{{ .Content }}<|im_end|>
{{ end }}<|im_start|>assistant
"""
PARAMETER stop "<|im_end|>"
PARAMETER stop "<|endoftext|>"
PARAMETER temperature 0.7
PARAMETER top_p 0.95
PARAMETER top_k 20
PARAMETER presence_penalty 1.5
PARAMETER repeat_penalty 1.0
PARAMETER num_ctx 32768
Build and run:
ollama create qwopus-kat-35b -f Modelfile
ollama run qwopus-kat-35b
</details>
<details>
<summary><b>3. Agentic IDE Setup (Cline / Roo-Code / Continue.dev)</b></summary>
<br>
Point your IDE agent extension directly to your local llama-server:
{
"apiProvider": "openai",
"apiBaseUrl": "http://localhost:8080/v1",
"apiKey": "local",
"modelId": "Qwopus-KAT-Coder-35B-Merged-APEX-Quality.gguf",
"contextWindow": 32768,
"maxTokens": 4096
}
</details>
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
Recommended Sampling Hyperparameters
> Important: Do not use generic ultra-low temperatures (0.1–0.2) or high repetition penalties (>1.05) on Qwen3.6-35B-A3B MoE models. Doing so causes expert activation collapse and repetitive looping. Use the tested settings below:
temperature: 0.7 # Optimal range is 0.6–0.9 for reasoning + code synthesis
top_p: 0.95
top_k: 20
min_p: 0.00 # Keep at 0.00 to avoid over-filtering sparse expert activations
presence_penalty: 1.5 # Keeps multi-step logic focused and avoids reasoning drift
repeat_penalty: 1.0 # Stay at 1.0 (do not exceed 1.05)
context_window: 32768
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
Prompt Template & Chat Template Options
⭐ Recommended: Improved Chat Template for Qwen 3.x
For optimal tool-calling accuracy, strict schema adherence, and multi-turn stability, use the standalone Improved Chat Template for Qwen 3.x:
- Repository:
OliviaRossi/Improved-Chat-Template-for-Qwen-3.x - Key Enhancements:
- Fixes tool-call XML/JSON schema formatting bugs in complex multi-file coding workflows.
- Eliminates reasoning/thinking tag leakage into assistant responses.
- Properly formats tool execution returns and error traces across long multi-turn sessions.
To use it with llama-server or vLLM:
# Download template
curl -LO https://huggingface.co/OliviaRossi/Improved-Chat-Template-for-Qwen-3.x/raw/main/chat_template.jinja
# Pass directly to llama-server
llama-server --chat-template-file chat_template.jinja ...
---
Default Template Structure (ChatML)
<|im_start|>system
You are an autonomous AI coding assistant. You analyze complex codebases, write clean and efficient code, debug issues systematically, and produce complete, working implementations.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
Acknowledgments & Lineage
| Role | Model / Resource | Link |
| :--- | :--- | :--- |
| Ultimate base architecture | Qwen3.6-35B-A3B (Alibaba Cloud) | — |
| Qwopus base fine-tune | Jackrong/Qwopus3.6-35B-A3B-Coder | huggingface.co |
| KAT-Coder base fine-tune | Kwaipilot/KAT-Coder-V2.5-Dev | huggingface.co |
| Merged base weights | OliviaRossi/Qwopus-KAT-Coder-35B-Merged | huggingface.co |
| GGUF Quantizations | OliviaRossi/Qwopus-KAT-Coder-35B-Merged-GGUF | huggingface.co |
| Recommended Chat Template | OliviaRossi/Improved-Chat-Template-for-Qwen-3.x | huggingface.co |
| Quantization Toolchain | llama.cpp & apex-quant | llama.cpp · apex-quant |
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
Citation
@misc{qwopus-kat-coder-35b-merged,
title = {Qwopus-KAT-Coder-35B-Merged},
author = {OliviaRossi},
year = {2026},
note = {SLERP merge of Jackrong/Qwopus3.6-35B-A3B-Coder and Kwaipilot/KAT-Coder-V2.5-Dev with APEX-Quality MoE quantization},
url = {https://huggingface.co/OliviaRossi/Qwopus-KAT-Coder-35B-Merged-GGUF}
}
<p align="right"><a href="#readme-top">↑ back to top</a></p>
---
License
Released under the Apache 2.0 license. Usage is also subject to the licenses of the underlying parent models.
<div align="center">
Built for developers who want a private, autonomous coding agent running locally on consumer hardware.
</div>
Run OliviaRossi/Qwopus-KAT-Coder-35B-Merged-GGUF with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models