Qwen3.8.6.5-27B-SymmTSV-Agentic-Coder
Symmetrical model merge for coding
1 min read
This model is a symmetrical merge of three baseline Qwen versions (3.5, 3.6, and 3.8 at 27B size) tailored for coding and agentic tasks. It is built on the SYMM-TSV-CoM method featuring per-token goal routing and weight orthogonalization. This approach combines the strengths of all three donors without hierarchical bias. The solution is ideal for developers and researchers who need a powerful tool for code generation and autonomous agents.
Key features
- Symmetrical fusion of three models with zero hierarchy drift
- Per-token goal routing using the Weiszfeld geometric median
- Löwdin orthogonalization to eliminate weight interference
- Improved negative log-likelihood (NLL/PPL) benchmark scores
What you can do
- Integrate into development environments for code generation and autocomplete
- Use as a backbone for building advanced coding agents
- Analyze complex software structures and optimize algorithms
Example
An example of using the model as a coding assistant to solve complex architectural tasks and write Python or C++ code using context across multiple files.
Read more
DetailsNLL/PPL Benchmark (Merged): 1.5158 (4.553)
Base models: Qwen3.5-27B, Qwen3.6-27B, Qwen3.8-27B
Parameters: 27.36 billion
MTP Policy: strip










Comments