chetvil https://chetvil.com/ Ilya Chatsviorkin on running open-weight models locally on an NVIDIA DGX Spark: DeepSWE agent benchmarks, serving recipes, throughput and cost. en Thu, 08 Oct 2026 00:00:00 GMT Best LLM for a single DGX Spark: Qwen3.8 Flash Next DeepSWE benchmark results https://chetvil.com/qwen3-8-flash-next-dgx-spark-deepswe/ https://chetvil.com/qwen3-8-flash-next-dgx-spark-deepswe/ Thu, 08 Oct 2026 00:00:00 GMT On DeepSWE it solves more tasks than DeepSeek V4 Flash and runs four of them at once on a single 128 GB Spark. Results from all 113 tasks. Can a fast decision model tell a correct patch from a broken one? https://chetvil.com/can-a-decision-model-judge-code-patches/ https://chetvil.com/can-a-decision-model-judge-code-patches/ Sun, 04 Oct 2026 00:00:00 GMT I asked TypeSafe's Jev to score 113 DeepSWE patches from my local DeepSeek run. Its scores did not separate passing patches from failing ones. Running DeepSeek V4 Flash across all of DeepSWE on one DGX Spark https://chetvil.com/deepseek-full-deepswe-run/ https://chetvil.com/deepseek-full-deepswe-run/ Sun, 13 Sep 2026 00:00:00 GMT A quantized open-weight model (EXL3 3-bpw) on a single 128 GB DGX Spark ran the entire DeepSWE 1.1 suite at max reasoning: 45% solved, with per-task results and failure analysis. Can a local model on a DGX Spark match a frontier model? A DeepSWE case study https://chetvil.com/deepseek-v4-flash-dgx-spark-vs-frontier/ https://chetvil.com/deepseek-v4-flash-dgx-spark-vs-frontier/ Mon, 07 Sep 2026 00:00:00 GMT Running DeepSeek V4 Flash on a single NVIDIA DGX Spark against GPT-5.6 Terra and Claude Sonnet 4.6 on one hard coding task. Initial Commit; https://chetvil.com/welcome/ https://chetvil.com/welcome/ Mon, 25 May 2020 00:00:00 GMT Why I started this blog, and how it's hosted.