Why this matters: small dense models that run locally while supporting agentic tool use and multimodal input are increasingly useful for on-device development and research. Qwopus3.5-9B-coder targets that niche by combining trace-inversion augmentation with real agent trace data to improve tool-calling reliability and stepwise reasoning for coding tasks.
Key Capabilities
- Fine-tuned for agentic coding and tool calling: trained with high-quality agent trajectories (lambda/hermes-agent-reasoning-traces) and trace inversion augmentation to expose deeper, stepwise reasoning signals — so what? the model is better calibrated to produce actionable tool calls and multi-step code workflows rather than only high-level suggestions.
- Local-friendly 9B dense (GGUF): designed to run in 8-bit on 16GB RAM devices, enabling faster local inference on laptops and small servers — so what? you can prototype agentic workflows and repository-level edits without large cloud GPUs.
- Multimodal + tool integrations: supports vision (requires mmproj.gguf alongside the main GGUF) and shows strong tool-call stability in the author's ToolCall-15 tests — so what? it’s practical for multimodal agent experiments that invoke external tools or terminals.
- Benchmark/usage signals: community metrics show non-trivial adoption (downloads: 6,462; likes: 69) and reported evaluation results (HermesAgent-20: 85; ToolCall-15: 100; BugFind-15: 79; SWE-bench Verified: 53.33), which suggest particular strength on agent/tool tasks relative to general software-engineering benchmarks.
Who it's for & trade-offs
Great fit if you: want a locally runnable 9B model for building coding agents, experimenting with tool-calling formats and multimodal prompts, or reproducing agent trace training methods. Look elsewhere if you: need rigorously validated production-grade generalist performance, strict safety/audit guarantees, or top-tier repository-level coding performance — the author flags this as an experimental community release and notes potential capability drift outside agentic programming scenarios. Also note license: Apache-2.0, and that vision support requires placing the provided mmproj.gguf file with the main model file. When deploying agent features, keep the model’s expected prompt/format (e.g., <think> tags) aligned with the training style to activate its tool-calling behavior.
