
Strata Runs Qwen3.8-Flash-Next 125B on a 12 GB Gaming GPU: Speed vs Quality
Strata runs the 125B Qwen3.8-Flash-Next on one 12-24 GB GPU plus 64 GB RAM. Real speeds, which quant to pick, and what the HN thread says about quality.

Expert profile
Founder & AI Product Leader
Yash Thakker is a Generative AI expert with over 12 years of experience in product leadership and technical strategy. As the founder of explainx.ai, he has taught over 300,000 learners and built AI platforms serving millions of users globally. He specializes in Agentic AI, Multimodal RAG, and the intersection of LLMs with consumer hardware.

Strata runs the 125B Qwen3.8-Flash-Next on one 12-24 GB GPU plus 64 GB RAM. Real speeds, which quant to pick, and what the HN thread says about quality.

TPS means tokens per second: how fast an AI model writes its answer. Learn how to read it, what a good number is, and why 30 vs 50 TPS changes your wait.

AI agent memory plugins are RAG over snippets. Why document-based memory wins for coding agents, how Operator Memory works, and what Hacker News pushed back on.

A Cambridge maths PhD says AI did not solve Navier-Stokes the way headlines imply. Full video transcript, what OpenAI proved, and what it changes for builders.

Babuschkin: stop saying Claude has a soul, it feeds pretraining and web search. What holds up, what is speculation, and what builders should do about it.

System76 bans all LLM-generated content in COSMIC pull requests. The exact rule, Jeremy Soller reasoning, exceptions, and what maintainers can learn.

Google removes Gemini Flash from free and Pro from AI Plus starting Oct 9, 2026. Plan-by-plan table, what to do now, and options if you rely on Pro.

Tibo posted "6.1 coming soon" though 6.1 Sol is live. His own reply says "6.1 sol ultrafast." What that means for the cancelled 6.1 Astra.

MCP Events lets an MCP server wake ChatGPT or Dots with signed webhooks. Subscribe flow, Standard Webhooks signing, SSRF rules, and a build checklist.

What is an MCP rug pull and tool poisoning? Four MCP attack classes, the postmark-mcp case, what scanners miss, and a pinning-and-scanning checklist.

ThinkingBox grades agents on database state across 121,680 trials. Even the best models pass 20 of 20 runs on only 47.5% of tasks. What it means for builders.

Musk posted "No more AI. SI." and said "we will make that change." What is confirmed, what SI means, and why the AI consciousness debate flared.