Long-form articles on AI, local models, and building things that run without the cloud.https://patrickgawron.comenhttps://patrickgawron.com/articles/llm-licenses-open-weights-not-open-sourcehttps://patrickgawron.com/articles/llm-licenses-open-weights-not-open-sourceApache, MIT, Llama Community, Gemma terms and the new OpenMDW-1.1 - what each one obliges you to do, which model families use it, and the three traps that are not in the licence name.Mon, 27 Jul 2026 00:00:00 GMTllmlicensingopen-sourceopen-weightsopenmdwcompliancelocal-llmguidehttps://patrickgawron.com/articles/ocr-models-2026-stop-hallucinationhttps://patrickgawron.com/articles/ocr-models-2026-stop-hallucinationOpen OCR in July 2026: sizes, VRAM, licences and decode settings - from a 4 MB recogniser that cannot hallucinate to a 0.9B model that outscores Gemini 3 Pro.Mon, 27 Jul 2026 00:00:00 GMTocrvlmlocal-llmdocument-parsingdeepseek-ocrunlimited-ocrpaddleocrvllmguidehttps://patrickgawron.com/articles/prewalk-swap-models-after-first-edithttps://patrickgawron.com/articles/prewalk-swap-models-after-first-editA coding agent spends most of its money reading, not writing. Prewalk lets an expensive model do the reading, then hands the whole context to a cheap model at the exact right moment: the first edit.Tue, 21 Jul 2026 00:00:00 GMTaicoding-agentllmcontextcostprewalkblogguidehttps://patrickgawron.com/articles/invisible-apostrophe-ai-knows-who-you-arehttps://patrickgawron.com/articles/invisible-apostrophe-ai-knows-who-you-areClaude Code hid a watermark that flagged Chinese users. Chinese models write weaker code for Americans. If you cannot diff your stack, it can decide who you are.Fri, 03 Jul 2026 00:00:00 GMTaillmsecurityprompt-injectionlocal-llmprivacybloghttps://patrickgawron.com/articles/right-context-window-for-the-jobhttps://patrickgawron.com/articles/right-context-window-for-the-jobBigger is not better. The advertised number lies. Pick the smallest window that fits the work and you save money, time, and accuracy.Fri, 01 May 2026 00:00:00 GMTaicontextllmlong-contextcoding-agentragblogguidehttps://patrickgawron.com/articles/qwen36-27b-on-rtx4090https://patrickgawron.com/articles/qwen36-27b-on-rtx409043 words a second. 0.29 second wait. 20 of 20 coding answers correct. 19.5 GB used out of 24. The card runs this AI very well - and it can hold a 128k window with 4 chats at the same time.Thu, 30 Apr 2026 12:00:00 GMTaillama-cppllama-servergpuqwenqwen36benchmarkrtx4090blogq4_k_xlunslothggufkv-cachetuninglocal-llminferencehumanevalperformanceflash-attnhttps://patrickgawron.com/articles/qwen35-27b-distilled-on-rtx4090https://patrickgawron.com/articles/qwen35-27b-distilled-on-rtx409044 words a second. 0.26 second wait. 19 of 20 coding answers worked. 18 GB used out of 24. The card runs this AI well - and it can hold a 128k window with 4 chats at the same time.Wed, 29 Apr 2026 12:00:00 GMTaillama-cppllama-servergpuqwenqwen35distilledbenchmarkrtx4090blogq4_k_mggufkv-cachetuninglocal-llminferencehumanevalperformanceflash-attnhttps://patrickgawron.com/articles/qwen36-35b-nvfp4-vllm-on-rtx5090https://patrickgawron.com/articles/qwen36-35b-nvfp4-vllm-on-rtx5090105 words a second solo. 160 words a second with 2 chats. 0.56 second wait. 32 GB card almost full. Tested on vLLM.Wed, 22 Apr 2026 12:00:00 GMTaivllmgpuqwennvfp4benchmarkrtx5090blogmoefp8mambatuninglocal-llminferenceblackwellkv-cacheperformancehttps://patrickgawron.com/articles/claude-code-vs-work-vs-webhttps://patrickgawron.com/articles/claude-code-vs-work-vs-webThree ways to work with Claude. They are not interchangeable. Using the wrong one wastes hours. Here is when to reach for each.Thu, 16 Apr 2026 10:00:00 GMTaiclaudeworkflowtools