https://aistack.imec-int.com/insights Independent, reproducible measurements of AI workloads on next-generation silicon, from imec's aistack. en https://aistack.imec-int.com/blog/quantization https://aistack.imec-int.com/blog/quantization Thu, 08 Oct 2026 00:00:00 GMT Does squeezing a model into fewer bits hurt a coding agent? FP8, INT4 AWQ and NVFP4 tested with Qwen3.8-27B and GLM5.3-Flash on agentic SWE benchmarks. https://aistack.imec-int.com/blog/agents-peeking https://aistack.imec-int.com/blog/agents-peeking Tue, 15 Sep 2026 00:00:00 GMT Coding agents found the answers in the sandbox's git history. After we closed the git, internet and recall routes, resolve rates fell from 75–90%+ to 40–53%. https://aistack.imec-int.com/blog/astra-vs-fable-update https://aistack.imec-int.com/blog/astra-vs-fable-update Tue, 15 Sep 2026 00:00:00 GMT Astra solves 34 of 64 SWE-bench Pro tasks for €53, Fable 51 for €121. Per solved task that is about €1.6 vs €2.4: Astra is cheaper because it is short. https://aistack.imec-int.com/blog/harness-cost https://aistack.imec-int.com/blog/harness-cost Tue, 01 Sep 2026 00:00:00 GMT Three coding harnesses, two models, 64 SWE-Bench Pro tasks. Harness choice barely moves resolve rate, but harness and model together can double your token bill. https://aistack.imec-int.com/blog/gpu-waiting https://aistack.imec-int.com/blog/gpu-waiting Mon, 27 Jul 2026 00:00:00 GMT A back-of-the-napkin guide to LLM inference on GPUs: the roofline model, prefill vs decode, the KV cache, and throughput vs per-user latency. https://aistack.imec-int.com/blog/gpu-self-hosting https://aistack.imec-int.com/blog/gpu-self-hosting Thu, 23 Jul 2026 00:00:00 GMT The first article in a series digging into the real cost and technicalities of owning your own AI coding agents — on hardware you actually control.