https://stondo.github.io/tags/deepseek-v4.1-flash/Recent content in Deepseek-V4.1-Flash on Bits-EntangledHugo -- 0.162.1en-usSat, 12 Sep 2026 14:15:00 +0000https://stondo.github.io/posts/dsv41-flash-complete-recipe-two-rtx-pro-6000/Sat, 12 Sep 2026 14:15:00 +0000https://stondo.github.io/posts/dsv41-flash-complete-recipe-two-rtx-pro-6000/The companion runbook to the V4.1-Flash deployment and debugging posts: the exact serve script, container invocation, systemd wiring, and generation defaults the lane runs in production right now — with the reason every flag is what it is, and the numbers to expect from each concession.https://stondo.github.io/posts/dsv41-flash-word-salad-top-p-opencode/Sat, 12 Sep 2026 13:00:00 +0000https://stondo.github.io/posts/dsv41-flash-word-salad-top-p-opencode/The new V4.1-Flash lane benchmarked clean for hours, then a user report: it starts producing garbage after a while, and only in some CLIs. The hunt ran through four test batteries that found nothing, a tcpdump that captured nothing, a reverse proxy that caught the actual payloads — and ended at a two-line JSON file fixing a sampling default nobody had set.https://stondo.github.io/posts/deepseek-v41-flash-552b-exl3-two-rtx-pro-6000/Sat, 12 Sep 2026 10:30:00 +0000https://stondo.github.io/posts/deepseek-v41-flash-552b-exl3-two-rtx-pro-6000/DeepSeek dropped V4.1-Flash on a Thursday: 552B backbone parameters plus a 196B Engram memory, activating 8-16B per token. Two days later it runs on my two RTX PRO 6000s at EXL3 2bpw — because a quant existed for exactly this hardware. This is the fit analysis, the deployment, and the honest gap between my 123 GB of host RAM and the 113 tok/s headline.