The ISO includes the proprietary NVIDIA stack so NVIDIA desktops and laptops have a working path on first boot. Non-NVIDIA systems remove the stack after first boot to reclaim disk. Hybrid laptops can still vary by vendor firmware. PRIME offload is the intended path for discrete-GPU apps.
+NVIDIA is an explicit, simulate-first path. The proprietary stack is not auto-installed and is not included then removed after first boot. The workflow refuses removals and can create Phoenix Points on Btrfs. Hybrid laptops and physical accelerator performance still need hardware-specific validation. After a user-installed NVIDIA driver, PRIME offload is the intended path for discrete-GPU apps.
__NV_PRIME_RENDER_OFFLOAD=1 __GLX_VENDOR_LIBRARY_NAME=nvidia blenderAnyone who needs Secure Boot signing, enterprise support, a Debian-stable base, or a zero-surprise production workstation should wait and follow the changelog.
The proprietary NVIDIA stack is included and hybrid laptops use PRIME offload, but laptop firmware varies. Read the hardware notes and known issues before installing.
+NVIDIA setup is an explicit, simulate-first workflow that refuses removals and can create Phoenix Points on Btrfs. It is not auto-installed and is not included then removed after first boot. Hybrid laptops and physical accelerator performance still need hardware-specific validation. Read the hardware notes and known issues before installing.
The local chat stack is designed to run on your machine. If you download models, those downloads come from the model host you choose. Shadowfetch does not add an account requirement or telemetry daemon to use the OS.
@@ -1501,6 +1502,126 @@ function faqPage() { }); } +function benchmarksPage() { + return shell({ + title: "Historical 16GB local LLM benchmark — Shadowfetch Linux", + canonical: "/linux/benchmarks", + description: "A dated August 2026 Ollama benchmark on one NVIDIA RTX 5060 Ti: measured tokens per second, VRAM footprint, method, and explicit limits. It is not a Shadowfetch Linux 3.5.0 runtime claim.", + body: ` +Dated August 2026. This is the old RTX 5060 Ti / ollama pull run. It is not a Shadowfetch Linux 3.5.0 runtime claim, and no 3.5.0 benchmark results are published here.
Speed is measured on the 16 GB card above. Whether a model fits is mostly its memory footprint, which travels between cards far better than its speed does.
+ +| Model | +Params | +Quant | +VRAM used | +On GPU | +Context | +Generation | +Prompt | +Cold load | +
|---|---|---|---|---|---|---|---|---|
Gemma 3 · 1Bollama pull gemma3:1b |
+ 999.89M | +Q4_K_M | +1.3 GB | +100% GPU | +32768 | +263 tok/s | +7931 tok/s | +2.61s | +
Qwen2.5 · 3B Instructollama pull qwen2.5:3b-instruct |
+ 3.1B | +Q4_K_M | +3.1 GB | +100% GPU | +32768 | +175 tok/s | +9768 tok/s | +2.66s | +
Gemma 3 · 4Bollama pull gemma3:4b |
+ 4.3B | +Q4_K_M | +4.6 GB | +100% GPU | +131072 | +120 tok/s | +3697 tok/s | +3.14s | +
Llama 3.2 · 3Bollama pull llama3.2:3b |
+ 3.2B | +Q4_K_M | +4.8 GB | +100% GPU | +131072 | +174 tok/s | +8985 tok/s | +2.92s | +
Qwen2.5 · 7B Instructollama pull qwen2.5:7b-instruct |
+ 7.6B | +Q4_K_M | +6.3 GB | +100% GPU | +32768 | +88 tok/s | +4960 tok/s | +8.55s | +
Qwen2.5 · 14Bollama pull qwen2.5:14b |
+ 14.8B | +Q4_K_M | +13 GB | +100% GPU | +32768 | +45 tok/s | +2548 tok/s | +9.97s | +
For the current release boundary, see local AI and coding agents. Buzz owns current model selection and download after confirmation.
+ +ollama ps reports for the resident model — the model's real footprint, and whether all of it sits on the GPU or spills to the CPU (which is the moment speed falls off a cliff).ollama show — the model file's own metadata, not typed by us. Everything shown is Q4_K_M-class weights, the quant most people actually run.