 ── Step 5 of 5  Results ────────────────────────────────────────────────

 server · context L 32k
 room      graphics 86.0 GB for Ollama (4 x 22.5 GB cards minus reserve)
           memory 86.3 GB for Ollama (94.3 GB minus reserve)
           graphics up to 82.8 GB for vLLM (4 x 22.5 GB cards, 0.92 of each)
           disk unknown (modelroom render reads no machine live)
 runtime   Ollama · vLLM (from its profile)
 One section per runtime: alternatives on the same cards, not both at once.

 with Ollama
 #    Model                   Package                Fit          Speed         Memory    Download
 ───────────────────────────────────────────────────────────────────────────────────────────────────
 1    Qwen3-8B                unsloth · Q4_K_M       ample        –             50.7 GB   4.7 GB
 shown     1 of 1 packages
 memory    #1 fits into graphics memory, spread over the cards, budget 86.0 GB
           (cards minus the reserve on each, from the profile)
           memory reserve 8.0 GB is set in modelroom.toml;
           computed from this machine it would be 16.0 GB
 speed     nothing measured in the rows shown · say Yes in step 4 to measure an installed package
 key       Fit is memory only: the package's need against a pool · ample: up to 60 % of the pool
           fits: up to 85 % · tight: up to 98 % · no fit: more
           the pool is what the mode names, minus its reserve (per card) · gpu: one card
           gpu_split: all cards together · cpu_gpu, cpu (RAM): the system memory
           system memory keeps a quarter as reserve, 2 to 16 GB, unless modelroom.toml sets it
           one memory shared by system and card counts as gpu, a quarter kept, no upper limit
           off the card, and for a fit from size the highest class is fits
           best: the least compression of the same model that fits
           modelroom rates no model family against another
           Speed is measured, see its column.

 with vLLM
 #    Model                   Package                Fit          Speed         Memory    Download
 ───────────────────────────────────────────────────────────────────────────────────────────────────
 2    Qwen3-8B                RedHatAI · AWQ         fits         –             56.7 GB   5.7 GB
 shown     1 of 1 packages
 memory    #2 vLLM on 4 cards: 0.92 of each, runtime 1.50 GB per card (one recording)
 key       Fit is memory only: how many requests of this context the cards hold at once
           ample: twice the requests this ranking assumes or more
           fits: that many, below twice them · tight: at least one, fewer than them
           no fit: not one · the pool is 0.92 of each card vLLM uses, 1.50 GB of it its runtime
           gpu: one card
           gpu_split: 2, 4 or 8 equal cards of the largest size, as the head count allows
           for a fit from size the highest class is fits
           best: the least compression of the same model that fits
           modelroom rates no model family against another
           Speed is measured, see its column.
 Quality of these models for your task: https://artificialanalysis.ai
