<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom" xml:lang="en">
<title>Blog - Blixild</title>
<subtitle>Posts on computer vision and GPU computing: measurement and visual inspection, machine monitoring, live camera systems, AI hosting and the hardware we rent.</subtitle>
<link rel="self" type="application/atom+xml" href="https://blixild.com/feed.xml" />
<link rel="alternate" type="text/html" href="https://blixild.com/blog.html" />
<id>https://blixild.com/feed.xml</id>
<updated>2026-11-22T00:00:00Z</updated>
<author><name>Jacek Garbuliński</name></author>
<rights>© 2026 Blixild</rights>
<entry>
<title>How long does Qwen3.8-27B take with a long document?</title>
<link rel="alternate" type="text/html" href="https://blixild.com/qwen3-8-27b-long-document.html" />
<id>https://blixild.com/qwen3-8-27b-long-document.html</id>
<published>2026-11-21T00:00:00Z</published>
<updated>2026-11-22T00:00:00Z</updated>
<summary>Qwen3.8-27B on one H200 took 52 to 55 seconds to first answer about a long document of 260,000 tokens. Later questions mostly came back in under 2 seconds.</summary>
</entry>
<entry>
<title>An EU LLM API with zero data retention, now in beta</title>
<link rel="alternate" type="text/html" href="https://blixild.com/eu-llm-api-blixild-ai-beta.html" />
<id>https://blixild.com/eu-llm-api-blixild-ai-beta.html</id>
<published>2026-11-20T00:00:00Z</published>
<updated>2026-11-22T00:00:00Z</updated>
<summary>Blixild AI, our EU LLM API, is in beta. Five open models answer from GPUs in Poland with zero data retention, from $0.05 per million tokens.</summary>
</entry>
<entry>
<title>How the basal LLM saves time and compute on text decisions</title>
<link rel="alternate" type="text/html" href="https://blixild.com/basal-llm-save-time-compute.html" />
<id>https://blixild.com/basal-llm-save-time-compute.html</id>
<published>2026-11-20T00:00:00Z</published>
<updated>2026-11-22T00:00:00Z</updated>
<summary>The basal LLM answered questions about texts 7 to 11 times faster than a 27B generative model on the same GPU, and got 35 of 38 right against 38.</summary>
</entry>
<entry>
<title>How to use Bielik 11B v3.1 through an API</title>
<link rel="alternate" type="text/html" href="https://blixild.com/bielik-api-guide.html" />
<id>https://blixild.com/bielik-api-guide.html</id>
<published>2026-11-20T00:00:00Z</published>
<updated>2026-11-22T00:00:00Z</updated>
<summary>In our tests through an API, Bielik 11B v3.1 placed a correct order as an agent every time, in 2 to 3 seconds, when its instructions set out each step.</summary>
</entry>
<entry>
<title>How to try Bielik 11B v3.1 on Blixild AI without code</title>
<link rel="alternate" type="text/html" href="https://blixild.com/bielik-11b-v3-1-how-to-try.html" />
<id>https://blixild.com/bielik-11b-v3-1-how-to-try.html</id>
<published>2026-11-19T00:00:00Z</published>
<updated>2026-11-22T00:00:00Z</updated>
<summary>Bielik 11B v3.1 answers through the Blixild AI API from a graphics card in Poland. A short reply costs about $0.0003, and trying it takes no code.</summary>
</entry>
<entry>
<title>How to build a self-hosted LLM API, from GPU to customer</title>
<link rel="alternate" type="text/html" href="https://blixild.com/self-hosted-llm-api-tech-stack.html" />
<id>https://blixild.com/self-hosted-llm-api-tech-stack.html</id>
<published>2026-11-19T00:00:00Z</published>
<updated>2026-11-22T00:00:00Z</updated>
<summary>A self-hosted LLM API lets a model on your own GPU serve customers, each with a key and a price per token, while the GPU has no address on the internet.</summary>
</entry>
<entry>
<title>How one RTX 3090 serves Arete 27B through an API</title>
<link rel="alternate" type="text/html" href="https://blixild.com/arete-27b-api-rtx-3090.html" />
<id>https://blixild.com/arete-27b-api-rtx-3090.html</id>
<published>2026-11-19T00:00:00Z</published>
<updated>2026-11-22T00:00:00Z</updated>
<summary>When we looked, no API offered Arete 27B. Since 18 November, one RTX 3090 in Poland serves it through Blixild AI, paid per token.</summary>
</entry>
<entry>
<title>How a camera tracks the moving parts of a machine</title>
<link rel="alternate" type="text/html" href="https://blixild.com/how-a-camera-tracks-machine-parts.html" />
<id>https://blixild.com/how-a-camera-tracks-machine-parts.html</id>
<published>2026-11-16T00:00:00Z</published>
<updated>2026-11-22T00:00:00Z</updated>
<summary>In machine monitoring, a camera can follow each moving part without training data. Its movements add up to running time, idle time and stops per shift.</summary>
</entry>
<entry>
<title>Just 13 minutes a shift can pay for machine monitoring</title>
<link rel="alternate" type="text/html" href="https://blixild.com/machine-monitoring-payback-2-percent-oee.html" />
<id>https://blixild.com/machine-monitoring-payback-2-percent-oee.html</id>
<published>2026-10-22T00:00:00Z</published>
<updated>2026-11-16T00:00:00Z</updated>
<summary>At a published 550 zł a month, machine monitoring of an old boring mill pays back from an OEE improvement of just over 2 points: 12.2 min a shift.</summary>
</entry>
<entry>
<title>Machine monitoring and downtime tracking without a PLC</title>
<link rel="alternate" type="text/html" href="https://blixild.com/machine-downtime-tracking-without-plc.html" />
<id>https://blixild.com/machine-downtime-tracking-without-plc.html</id>
<published>2026-10-19T00:00:00Z</published>
<updated>2026-11-22T00:00:00Z</updated>
<summary>For machine monitoring without a PLC, a current sensor or a stack light records when a machine runs. Only a camera also records each stop on video.</summary>
</entry>
<entry>
<title>NVIDIA H200 server available for rent</title>
<link rel="alternate" type="text/html" href="https://blixild.com/first-post.html" />
<id>https://blixild.com/first-post.html</id>
<published>2026-08-11T00:00:00Z</published>
<updated>2026-10-12T00:00:00Z</updated>
<summary>A server with an NVIDIA H200 (141 GB HBM3e) is available for rent. It serves one client at a time, and billing covers actual usage.</summary>
</entry>
</feed>
