New Capabilities
By deepening the model architecture and training the model with bi-directional logical reasoning, the model unlocks advanced cognitive and logical processing flows. It evaluates problems from multiple angles and revisits context iteratively, making it exceptionally powerful for complex problem-solving, multi-step logical deductions, and operating within autonomous agent frameworks.
This is a reasoning-heavy model. Please set your thinking budget to at least 8,192 tokens, and 16,384 tokens if working on a complex or difficult agentic, math, or coding workload.
Multi-Domain Execution Gates
By implementing Dynamic KL Divergence Masking and unlearning conflicting SFT ChatML tags, this model now natively supports complex Execution Gates. It intelligently identifies environments via XML tags and adapts its reasoning logic (<step> vs <think>) to perfectly match the underlying tool or language.
By employing Consistency-Gated GRPO during the RL phase, we aligned the V1 models sophisticated logical reasoning with the output by rewarding not just correct reasoning, but forcing the model to match the reasoning with correct code execution.
Core Verified Domains:
- C++ (ThreadSanitizer): Employs rigorous
<step> logic to generate data-race-free code (e.g. Meyers' Singleton pattern) directly into ````cpp` blocks.
- SWI-Prolog (Prover): Natively drops into
<think> traces to evaluate logical relations and construct valid ````prolog` mathematical proofs.
- Python (SWE-bench): Excels at strict PEP8 software engineering tasks, routing logic through precise
<step>-based analytical breakdowns.
- Agentic JSON: Bypasses standard logical traces entirely when strictly constrained, flawlessly generating pure, valid
<tool_call> JSON objects.
This is a reasoning-heavy model. Please set your generation limit to at least 8,192 tokens when evaluating C++ or SWE-bench workloads.
