Ollama Cloud Launches US/EU Zero-Data-Retention Hosting for DeepSeek-V4.1-Flash-1017 with Claude Code CLI, While SemiAnalysis Reveals GLM-5.3 Was Built 100% via Post-Training

Core Background & Industry Pain Points

As engineering teams integrate frontier open-weights coding models such as DeepSeek-V4.1-Flash-1017 and GLM-5.3 into terminal coding agents like Claude Code and OpenCode, two major questions dominate adoption: First, how can enterprise developers run open-weights models in the cloud with strict US/EU data residency, zero data retention, and zero training on proprietary code? Second, as pretraining compute costs escalate, can frontier labs achieve generational leaps in cybersecurity and agentic engineering purely through GPU-efficient post-training without pretraining a new base model from scratch?

Architectural Highlights & Core Mechanisms

Today's updates from Ollama (@ollama) and SemiAnalysis (@SemiAnalysis_) address both cloud privacy infrastructure and post-training scaling:

• Ollama Cloud US/EU Zero-Data-Retention Hosting & Native Agent Launchers: @ollama rolled out US- and EU-hosted Ollama Cloud with an unconditional always-zero-data-retention and no-training policy, making deepseek-v4.1-flash:1017-cloud immediately accessible via native ollama launch wrappers for Claude Code and OpenCode. Concurrently, Ollama v0.40.6 landed on GitHub with q2_0 GGUF tensor and file type support, clef cloud integration, and fixes for orphaned runner processes on Windows.
• SemiAnalysis Reveals GLM-5.3 Was Built 100% via Post-Training: In a newly published engineering clip (2111596468528087078), @SemiAnalysis_ highlighted that GLM-5.3 was built entirely by post-training the previous GLM-5.2 model, demonstrating that targeted post-training and multi-environment reinforcement learning are far more GPU-efficient than pretraining from scratch.

Authoritative Benchmarks & Empirical Comparisons

• Night-and-Day Leap in Cybersecurity Capabilities (GLM-5.3 vs GLM-5.2): According to @SemiAnalysis_, despite starting from the previous base checkpoint, GLM-5.3's intensive post-training pipeline produced a "night and day difference" in autonomous cybersecurity and system engineering evaluations compared to GLM-5.2 at a fraction of full pretraining GPU hours.
• Agentic Efficiency on Ollama Cloud: Running deepseek-v4.1-flash:1017-cloud brings 77.0% DeepSWE v1.1 and 42.3% Terminal-Bench 4.0 accuracy directly into Claude Code and OpenCode sessions with shorter output trajectories and lower cost per completed task.

Developer Practical Guide & Quickstart

After updating to Ollama v0.40.6, developers can launch zero-retention cloud-backed coding agents directly from their terminal:

• Interactive CLI: ollama run deepseek-v4.1-flash:1017-cloud
• Launch Claude Code: ollama launch claude --model deepseek-v4.1-flash:1017-cloud
• Launch OpenCode: ollama launch opencode --model deepseek-v4.1-flash:1017-cloud