32 GB RAM laptop: Qwen3.8-27B, every weight resident

MSI Prestige 16 AI+ C3MTG · Core Ultra X7 358H · 32 GB

One captured run per system

A 27B dense model on the same 32 GB laptop, every weight resident in memory on both systems. Both captures answer a merge sort in C; on Cosmos it's the first answer in a fresh chat. Thinking was set to medium on Cosmos and low on llama.cpp.

Laptop: MSI Prestige 16 AI+ C3MTG
CPU: Intel Core Ultra X7 358H
GPU: Intel Arc B390 · Xe3 (integrated)
Memory: 32 GB, shared by CPU, GPU and NPU

Left, kore status mem on the Core Ultra X7 358H laptop. Right, the Cosmos chat running QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4 at Effort: M…, with nothing above the user's message, Write an example of merge sort in C. Below it, Thought for 4 s and the start of the answer: a C listing with #include <stdio.h> and <stdlib.h>, merge, merge_sort and the start of main. — The request · CosmosMSI Prestige 16 AI+ C3MTG · Core Ultra X7 358H · 32 GB
The same Cosmos chat as the answer began: the prompt is its first message.

24.0tok/s

Decode speed on this laptop · 4.4× the speed of Ubuntu + llama.cpp on the same machine

Cosmos: 24.0tok/s
Ubuntu + llama.cpp: 5.44tok/s

Readouts, as shown on screen

Cosmos

586 tokens at 24.0 tok/s · context 2124 of 180224 · 5.38 tokens a step

Context 2,124 / 180,224 · 5.38 tokens per decode step · Medium thinking

Ubuntu + llama.cpp

638 tokens · 1min 57s · 5.44 t/s

638 tokens in 1 min 57 s · Low thinking

The thinking setting sets how long the model thinks before it answers; each figure is the decode rate of the answer itself. llama.cpp's is the fastest answer captured in its session, and its average across all turns read 3.7 t/s.

What each system ran

Weights: BothQwen3.8-27B · QUASAR · MTP · NVFP4
System: CosmosCosmos on Kore; Ubuntu + llama.cppUbuntu, on the same laptop · version not recorded
Runtime: CosmosKore's inference engine · built-in chat; Ubuntu + llama.cppllama.cpp, an earlier build · version not recorded · web UI

Evidence

Benchmark capture card, Pair C, 32 GB dense model, Qwen3.8-27B QUASAR NVFP4, Cosmos. Left, kore status mem on the Core Ultra X7 358H laptop. Right, the Cosmos chat: the end of a C merge sort answer, with merge and merge_sort functions and a main that sorts 38, 27, 43, 3, 9, 82, 10, then notes on the approach and its time and space complexity. The status line 586 tokens at 24.0 tok/s, context 2124 of 180224, 5.38 tokens a step is highlighted, as is the effort setting, which reads Effort: M… for Medium. The card reads 24.0 tokens per second, C merge-sort response, Medium thinking. — Cosmos · 24.0 tok/sMSI Prestige 16 AI+ C3MTG · Core Ultra X7 358H · 32 GB

Benchmark capture card, Pair C, Ubuntu with llama.cpp. The llama.cpp web UI running Qwen3.8 27B QUASAR NVFP4 MTP after a C merge sort answer, with 638 tokens, 1 min 57 s and 5.44 t/s highlighted, and the Low thinking setting highlighted in the input bar. The card reads 5.44 tokens per second, C merge-sort response, Low thinking. — Ubuntu + llama.cpp · 5.44 tok/sMSI Prestige 16 AI+ C3MTG · Core Ultra X7 358H · 32 GB

How it was measured

Same machine: Every comparison ran on one laptop, with the same silicon, drive and model weights on each side. On metal, Cosmos and Ubuntu were booted in turn; a result that ran Cosmos as a guest says how the guest was set up.
Current baseline: Ubuntu 26.04.1 LTS on Linux 7.0.0-38 with the latest Intel Xe drivers, on both machines; llama.cpp 0.8, with native NVFP4 and expert-cache support; fio 3.41 with io_uring, direct I/O and registered buffers, as the storage result's commands show. Current releases, not old builds, with one exception: the 27B result's llama.cpp run used an earlier build whose version wasn't recorded, and its settings say so.
Captured runs: Each figure is one captured run, read off its screenshot — not an average of repeated runs.
Native harnesses: Each system is measured with its own tools: Cosmos's chat and pack bench; llama.cpp's web UI and fio on Ubuntu. Transfer sizes and scheduling are reported, not equalized.
Units: Storage in decimal GB/s (10⁹ bytes per second). Tokens per second as each runtime reports its own generation speed.
Setup excluded: Buffer allocation and registration happen before timing on both systems; Cosmos prints its setup time on a separate line.

More results

Download Cosmos