Auditable memory estimates — architecture inputs are pinned to official configs; runtime and OS reserves remain estimates.
Qwen/Qwen3.8-Flash-Next validation status ↗ — identity verified, but five unmodeled architecture components mean no trustworthy fit estimate yet. FitLLM refuses to guess, so that checkpoint is not offered below. The “Qwen 3.8 27B” option in the calculator is a different, dense model (Qwen/Qwen3.8-27B) that FitLLM does model; its verdict says nothing about Qwen/Qwen3.8-Flash-Next.
Powered by fitllm-engine (MIT, single file, zero deps, conformance-vector tested) ·
CLI: npx fitllm "GLM-4.7-Flash" --gpu 4090
Full calculator with context & KV-cache controls: fitllm.run · No ads · No login · Output never for sale.