---

base_model:

• openbmb/MiniCPM-V-4.7-35B-A3B
• ornith-ai/Ornith-1.5-35B-A3B
• Kwaipilot/KAT-Coder-V2.5-Dev
• Accio-Lab/occamy-1.0

base_model_relation: merge

pipeline_tag: text-generation

language:

• en

tags:

• coding
• agent
• conversational
• long-context
• terminal
• llama.cpp
• gguf
• apex

---

Qwen3.6-Qworus-35B-A3B-V2-APEX-GGUF

What is this?

This is a DELLA merge of the following finetuned & post-trained models:

• openbmb/MiniCPM-V-4.7-35B-A3B: The base model.
• Kwaipilot/KAT-Coder-V2.5-Dev: Contributes token-efficient coding and more programming language knowledge.
• Accio-Lab/occamy-1.0: Contributes general terminal use, strong agent capabilities.
• ornith-ai/Ornith-1.5-35B-A3B: Contributes strong reasoning, self-scaffolding & architecture, coding, planning.

The result is a hybrid model for coding, planning, designing, terminal work or just conversation.

Compared to the first 35B-A3B "Qworus" merge, this one uses MiniCPM-V 4.7 as the base instead of Ornith and uses a mixture of ratios that gives stronger reasoning. Exhibits more ability in return for spending more tokens on thinking and working through a problem.

Although most of this model's weights were merged with with a method that drops weights and uses sign election to resolve conflicts, more sensitive tensors such as the router gates, shared experts, and token embed & output, were protected by using a density of 1.0 and epsilon of 0.0 to disable sparsification.

---

Recipe:

merge_method: della
base_model: openbmb/MiniCPM-V-4.7-35B-A3B
tokenizer_source: openbmb/MiniCPM-V-4.7-35B-A3B
dtype: bfloat16
out_dtype: bfloat16

models:
  - model: Kwaipilot/KAT-Coder-V2.5-Dev
    parameters:
      weight: 0.45
  - model: Accio-Lab/occamy-1.0
    parameters:
      weight: 0.30
  - model: ornith-ai/Ornith-1.5-35B-A3B
    parameters:
      weight: 0.25

parameters:
  int8_mask: true
  normalize: true
  lambda: 1.05
  density:
    - value: 0.67
    - filter: ".*(embed_tokens|lm_head|norm\\.weight|mlp\\.gate|mlp\\.shared_expert|self_attn\\.[qkvo]_proj).*" # Protect weights sensitive to dropping
      value: 1.0
  epsilon:
    - value: 0.045
    - filter: ".*(embed_tokens|lm_head|norm\\.weight|mlp\\.gate|mlp\\.shared_expert|self_attn\\.[qkvo]_proj).*"
      value: 0.0

---

Quantization

Weights are provided quantized with APEX, which preserves sensitive tensors with higher precision in respect to how MoE models work. The importance matrix was calibrated on about 50% random English text snippets from multiple domains, and 50% tokens generated during real long-context agent sessions with KAT-Coder and Ornith-1.5 (control tokens such as <|im_start|> included).