Qwen 3.5 40B Models Trained on Claude Outputs

What It Is

A series of fine-tuned language models has emerged that combines Qwen 3.5’s 40-billion parameter architecture with training data generated by Claude Opus. Developer DavidAU released three variants, each targeting different use cases: a standard reasoning-focused version, an uncensored “Heretic” variant, and a “Rough House” edition that pushes boundaries further.

All three models share the same 40B parameter count and 1275 tensor structure, but differ in their training approaches and content filtering. The standard version at https://huggingface.co/DavidAU/Qwen3.5-40B-Claude-4.5-Opus-High-Reasoning-Thinking emphasizes logical reasoning tasks. The Heretic variant (https://huggingface.co/DavidAU/Qwen3.5-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking) removes content restrictions, while the Rough House model (https://huggingface.co/DavidAU/Qwen3.5-40B-RoughHouse-Claude-4.6-Opus-Polar-Deckard-Uncensored-Heretic-Thinking) represents the most unrestricted option.

The training methodology involves distillation - using outputs from Claude Opus to teach Qwen 3.5 new capabilities. This approach attempts to transfer Claude’s reasoning patterns and response quality to a model that can run locally.

Why It Matters

This development highlights an ongoing trend in AI: using proprietary model outputs to enhance open-source alternatives. Developers gain access to Claude-like reasoning capabilities without API costs or rate limits. Research teams can experiment with model behavior in ways that commercial services restrict.

The availability of GGUF quantized versions, prepared by Mradermacher’s team, democratizes access significantly. Quantization reduces memory requirements while preserving most model capabilities, meaning developers can run these models on consumer hardware rather than requiring enterprise infrastructure.

The three-tier approach addresses different deployment scenarios. Organizations with strict content policies might choose the standard version, while researchers exploring edge cases benefit from the uncensored variants. This flexibility matters for teams building specialized applications where one-size-fits-all content filtering creates problems.

However, this also raises questions about model distillation ethics and licensing. Training on proprietary model outputs exists in a gray area - Anthropic’s terms of service may not explicitly permit this use case. The community continues debating whether such practices constitute fair use or violate service agreements.

Getting Started

Developers can download these models directly from HuggingFace. For local inference, the quantized GGUF versions work with popular frameworks like llama.cpp:

# Clone llama.cpp and build git clone https://github.com/ggerganov/llama.cpp cd llama.cpp && make

# Download a quantized model (example: Q4_K_M variant)
# Then run inference
./main -m qwen3.5-40b-claude-q4_k_m.gguf -p "Explain quantum entanglement"

For Python-based workflows, libraries like transformers or ctransformers provide integration options. The models follow standard chat templates, making them drop-in replacements for existing Qwen implementations.

Teams should start with the standard reasoning variant to evaluate performance before exploring the uncensored versions. Testing on representative tasks helps determine whether the Claude-based training actually improves results for specific use cases.

Context

These models compete with other distillation efforts like Nous Research’s Hermes series and various Llama fine-tunes. The 40B parameter count positions them between smaller 7B-13B models (faster but less capable) and 70B+ variants (more powerful but resource-intensive).

Direct comparisons with Claude Opus remain difficult since the original model isn’t available for local deployment. Early community feedback suggests these variants capture some reasoning characteristics but don’t fully replicate Claude’s performance on complex tasks.

Limitations include potential training data biases inherited from Claude, unknown dataset composition, and the usual challenges of running large language models (hallucinations, context window constraints, computational requirements). The uncensored variants require careful deployment consideration since they lack safety guardrails present in commercial offerings.

Alternative approaches include training on synthetic data from multiple sources, using reinforcement learning from human feedback, or fine-tuning on domain-specific datasets. Each method involves different tradeoffs between capability, control, and resource requirements.