Most modern code agents struggle to combine long-horizon context, stable tool usage, and RL-driven behavior refinement in a single open-weight release. This release exposes a post-trained MOE code model that was specifically tuned to reduce pathological tool calls and repetition while improving pass rates on agentic coding benchmarks, making it useful for researchers and teams experimenting with agentic coding stacks.
What Sets It Apart
- Targeted post-training recipe: built on Qwen3.6-35B-A3B and post-trained via supervised fine-tuning (127K examples) followed by RL with safeguards (TITO, Truncated Importance Sampling, sandbox verifiers) to stabilize policy updates for agentic workflows.
- Concrete scale and architecture: a Mixture-of-Experts model with 35B total parameters and ~3B activated parameters, optimized for tool-using coding agents rather than multimodal inference.
- Measured improvements: reported SWE-bench Verified 69.40, SWE-bench Multilingual 63.00, PinchBench 93.43, and competitive Terminal-Bench/Scicode/KAT-Code-Bench scores versus several 30–35B peers.
- Long-context and deployment-ready: native support for up to 262,144 tokens (with YaRN/rope overrides for longer horizons) and explicit guidance for running via vLLM, SGLang, Transformers and KTransformers as text-only model.
Who It's For & Tradeoffs
Great fit if you are a researcher or engineering team who needs an open-weight agentic coding model to: iterate on tool-use policies, run agent-in-the-loop code synthesis evaluations, or benchmark agentic coding harnesses. The release is explicit about being text-only: the vision/multimodal tower is not included, so it is unsuitable if you need multimodal capabilities out of the box. Serving requires significant GPU resources (tensor parallel setups are recommended) and careful inference flags (language-model-only / rope overrides) to avoid initialization errors.
Where It Fits
Use this model when you want an open checkpoint that: (a) reproduces RL-stabilization techniques for agentic coding, (b) integrates with common inference stacks (vLLM/sglang/transformers), and (c) supports ultra-long contexts. Look elsewhere for lightweight on-device models, turnkey multimodal agents, or proprietary hosted LLM services.
