Initial SQL-only 1C adapter baseline

registry/model-cards/qwen3-coder-30b-a3b-instruct-q6_k.yaml

@@ -0,0 +1,38 @@
+id: qwen3-coder-30b-a3b-instruct-q6_k
+name: Qwen3 Coder 30B A3B Instruct GGUF Q6_K
+type: gguf-model
+task:
+  - 1c
+  - code
+  - agentic-coding
+  - repository-analysis
+  - tool-use
+language:
+  - ru
+  - en
+source: huggingface
+upstream_id: lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF
+upstream_url: https://huggingface.co/lmstudio-community/Qwen3-Coder-30B-A3B-Instruct-GGUF
+license: apache-2.0
+status: candidate
+created_at: 2026-06-19
+storage_path: /models/gguf/1c/qwen3-coder-30b-a3b-instruct-q6_k
+filename: Qwen3-Coder-30B-A3B-Instruct-Q6_K.gguf
+file_size_bytes: 25104724288
+format: gguf
+quantization: Q6_K
+context_length: 262144
+deployment_context_length: 8192
+vram_required_gb: null
+parameters: 30.5B total / 3.3B active
+base_model: Qwen/Qwen3-Coder-30B-A3B-Instruct
+adapter_for: null
+datasets: []
+eval_suites:
+  - plugins/1c/evals/smoke.yaml
+deployment:
+  target: docker-gpu.cin.su
+  runtime: llama.cpp
+  served_model_name: qwen3-coder-1c-q6
+  compose: core/deploy/docker-gpu/llama-cpp/compose.yaml
+notes: "High-quality Q6_K quant for local 1C/code experiments on a single RTX 4090. Expected to require partial CPU/RAM offload or a short context because the GGUF file is larger than 24 GiB VRAM. Test first at 8K context, then 16K if stable."