<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>vLLM Bangkok — Blog</title><description>Notes on vLLM, inference and the Thai AI community — from the team building it here.</description><link>https://vllmbangkok.com/</link><language>en-US</language><item><title>vLLM Bangkok Day: the creators are coming to Bangkok</title><link>https://vllmbangkok.com/en/blog/vllm-bangkok-day-announced/</link><guid isPermaLink="true">https://vllmbangkok.com/en/blog/vllm-bangkok-day-announced/</guid><description>On 6 September 2026 the people who build vLLM are coming to Bangkok for an afternoon of production inference talks, Thai deployment stories and networking. Free to attend. Here is what to expect.</description><pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate><category>event</category><category>community</category><author>vLLM Bangkok</author></item><item><title>DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation</title><link>https://vllmbangkok.com/en/blog/dspark/</link><guid isPermaLink="true">https://vllmbangkok.com/en/blog/dspark/</guid><description>A paper review of how DSpark improves speculative decoding. In this article, we will go through how DSpark is designed and why it can improve throughput with a higher accepted length.</description><pubDate>Thu, 05 Nov 2026 00:00:00 GMT</pubDate><category>vllm</category><category>inference</category><category>speculative-decoding</category><category>paper-review</category><author>Natthanan Bhukan</author></item><item><title>Serving the agentic workload: Simon Mo on what AI agents change in vLLM</title><link>https://vllmbangkok.com/en/blog/serving-the-agentic-workload/</link><guid isPermaLink="true">https://vllmbangkok.com/en/blog/serving-the-agentic-workload/</guid><description>A summary of Simon Mo&apos;s Runtime 2026 talk: why agent sessions that read a median 142k tokens to write 444 put the KV cache at the centre of serving, what vLLM does about it, and why 2026 is the year of hardware heterogeneity.</description><pubDate>Wed, 04 Nov 2026 00:00:00 GMT</pubDate><category>vllm</category><category>inference</category><category>agents</category><category>kv-cache</category><category>talk-summary</category><author>Natdhanai Praneenatthavee</author></item><item><title>What is vLLM, and why does throughput jump when you switch to it?</title><link>https://vllmbangkok.com/en/blog/what-is-vllm/</link><guid isPermaLink="true">https://vllmbangkok.com/en/blog/what-is-vllm/</guid><description>A practical primer on the two ideas that make vLLM fast — PagedAttention and continuous batching — and what they mean for the GPU bill of a team serving open-weights models in Thailand.</description><pubDate>Sat, 08 Aug 2026 00:00:00 GMT</pubDate><category>vllm</category><category>inference</category><category>primer</category><author>vLLM Bangkok</author></item></channel></rss>