NPC Agentic 7B: A Single-GPU QLoRA Recipe for a Laptop-Scale Conversational Model

We present NPC Agentic 7B, a 7-billion-parameter conversational model trained via QLoRA on a single NVIDIA A40 GPU and quantized for on-device inference. The shipped Q4_K_M GGUF is 4.68 GB; on consumer Apple Silicon it runs at approximately 25 tokens/second, and on an A40 with W4A16 GPTQ via vLLM it serves at approximately 76 single-stream tokens/second. The model is built from Qwen2.5-7B-Instruct fine-tuned for 56 hours of wall-clock training on a mixture of 83,587 reasoning-trace examples and 3,000 cohort-weighted identity examples. We measure the model against its base on GSM8K-100 and the Berkeley Function Calling Leaderboard (BFCL v4 single-turn and multi-turn categories). The training improves irrelevance detection on BFCL (+9.6 percentage points) and regresses BFCL function-calling on most other subsets (-2 pp single-turn, -8 pp multi-turn average, peak -13 pp on multi_turn_miss_func) and GSM8K (-55 pp). We isolate the function-calling regression to an over-reasoning failure mode: under a 4,096-token completion budget, v3's <think>-style reasoning consumes the budget before the model emits the required <tool_call> block. Practitioners deploying v3 should set max_new_tokens ≥ 8192 or use </think> as a stop sequence. The contribution is a documented end-to-end recipe (data, hyperparameters, label masking, post-training quantization) and an honest set of base-vs.-fine-tune benchmarks identifying what the recipe gains, what it costs, and the deployment configuration required to use the resulting model effectively. Four artifacts are released on Hugging Face under Apache-2.0: bf16 merged, LoRA adapter, GPTQ W4A16, GGUF (Q4_K_M / Q5_K_M / Q8_0).