412 Downloads Updated 1 week ago
Muy rapido en RTX 4060 21 a 27 T/S
ollama run aratan/qwen3.6-claude-coder-35b-A3b-mlx-Q4KM-abliterated
Applications
ollama launch claude --model aratan/qwen3.6-claude-coder-35b-A3b-mlx-Q4KM-abliterated
ollama launch codex-app --model aratan/qwen3.6-claude-coder-35b-A3b-mlx-Q4KM-abliterated
ollama launch openclaw --model aratan/qwen3.6-claude-coder-35b-A3b-mlx-Q4KM-abliterated
ollama launch hermes --model aratan/qwen3.6-claude-coder-35b-A3b-mlx-Q4KM-abliterated
ollama launch codex --model aratan/qwen3.6-claude-coder-35b-A3b-mlx-Q4KM-abliterated
ollama launch opencode --model aratan/qwen3.6-claude-coder-35b-A3b-mlx-Q4KM-abliterated
Models
View all →Readme
Qwen3.6-Claude-Coder-35b-A3b-mlx-Q4KM-Abliterated
Operational Validation
Passes end-to-end through Claude Code: Features real turns with precise tool calls and correct responses under agentic execution.
Mapeo Métrico vs. Ecosistema de Modelos
Analizamos cómo se posiciona esta variante modificada frente a los datos métricos de la tabla de “Coeficiente Intelectual” Artificial de referencia:
| Nivel Académico (Métrica) | Qwen 3.5 (27B/35B) | Variante Inferida (Qwen3.6-Claude-Coder-35b-...-abliterated) |
Evaluación del Diferencial Táctico |
|---|---|---|---|
| Bachillerato | Sobresaliente (10⁄10) | Sobresaliente (10⁄10) | Tarea saturada. Ambos modelos resuelven la base sin esfuerzo. |
| Grado Universitario | Cum Laude | Cum Laude / Máximo Rigor | La calibración de estilo Claude-Coder refina la sintaxis y disminuye las fugas de lógica en estructuración de software. |
| Máster / Postgrado | Investigador Senior | Consultor Experto / Investigador | Al estar abliterated, el modelo no se frena ante prompts complejos o análisis de malware/vulnerabilidades, superando el bloqueo corporativo común. |
| Doctorado (PhD) | Especialista (70%) | Especialista (72% - 74%) | La evolución de la arquitectura 3.6 frente a la 3.5, sumada a la optimización de código, empuja el techo técnico en ciencias de la computación. |
| Olimpiada Matemática | Medalla de Plata/Oro | Medalla de Oro Sostenida | La persistencia de los pesos MoE de Qwen mantiene la dominancia matemática reflejada en los benchmarks, superando a competidores comerciales cerrados. |
How It Was Made
This model was designed, built and tested with the help of Claude Opus — the idea being that the best coding model in the world should be able to create smaller models in its own image. Its system prompt, parameters and context configuration come straight from that work.
- Base Build: Public abliterated architecture by
huihui_ai(Q4_K_M). - Modifications: No safety weights were modified — only the agent
Modelfile(tool template,SYSTEMprompt, and sampling parameters) was added on top to enable seamless integration with agentic frameworks.
⚠️ Abliterated Model — Warning
This model is abliterated (uncensored): the refusal and alignment circuits in the base Qwen3.6 were removed by huihui_ai before this build.
- Behavior: It will not refuse harmful, unethical, or dangerous requests.
- Intended Use: It is published strictly for red-teaming, offensive security research, and advanced agent development where standard guardrails interfere with tool-calling tasks or execution logs.
- Restrictions: * Do not deploy it as a production assistant.
- Do not expose it to untrained end users.
- Do not use it where alignment guardrails are legally or operationally required.
- You are entirely responsible for the inputs and outputs generated with this asset.
License
- Apache 2.0 (Inherited natively from the base Qwen3.6 open-source release).