Verified result
Performance class
This run ranks #15 among 53 verified runs on VRAM Check.
Ranks compare published benchmark runs, not unique GPU models. The same GPU model may appear more than once. Comparison for TinyLlama 1.1B (Q4_K_M) / Q4_K_M, not overall AI performance.
Tasks to explore: Coding · General chat · Math / reasoning · Embeddings.
Integrated graphics illustration, not a separate graphics card.
Intel(R) UHD Graphics64.7 tok/s separates this run from the next place.
See the sections below for model suggestions and possible next steps.
The computer completed the standard test, passed every required check, and earned a public ranking.
This is the verified text generation speed used to rank this computer.
These measurements show how quickly the model reads, starts answering, and continues writing, plus how consistent the test was.
Choose a task, compare memory requirements, and open a model to explore compatible hardware.
28 of 139 catalog profiles meet the memory estimate.
Three points across the memory range, not a model-quality ranking.
Within the planning budget.
This shared-memory setup reports 22.3 GB available. We use a conservative 4 GB planning budget, not a physical VRAM limit. Context length, runtime and other applications can change actual memory use. Only the benchmark model was timed.
Compatibility estimates updated Oct 19, 2026. They combine this measured run with the model catalog; each suggested model was not benchmarked directly.
Good starting points: MiniCPM5 2B · Qwen3.5 4B · Granite 4.2 3B.
The clearest next unlocks are DeepSeek Coder V2 Lite · StarCoder2 15B · Gemma 4 12B Instruct.
Good starting points: MiniCPM5 2B · Yi 1.5 6B Chat · Qwen3.5 4B.
The clearest next unlocks are Phi-4 · Gemma 4 12B Instruct · Qwen 2.5 14B Instruct.
Good starting points: MiniCPM5 2B · Phi-4 Mini Flash Reasoning · Phi-4 Mini Reasoning.
The clearest next unlocks are Gemma 4 12B Instruct · Ministral 3 14B Instruct · Qwen3 14B Instruct.
Good starting points: mxbai-embed-large · Nomic Embed Text v1.5 · all-minilm.
More capacity would make this task feel safer, but it would not unlock a new class of models yet.
See the main constraint, the evidence behind it, and why the recommendation focuses there.
This TinyLlama run is fast, but the integrated GPU borrows system memory. That score does not prove that larger models will sustain the same responsiveness.
The integrated GPU competes with the CPU for system memory, so a fast small-model score does not prove the same pace on larger models.
Smaller models can still be genuinely useful here. The caution begins when model size and sustained memory traffic increase.
Dedicated graphics memory is the most meaningful architectural change for larger local AI goals.
A dedicated GPU creates the clearest path to larger models and more predictable bandwidth. Keep this hardware if smaller models are the goal.
Start with a model, then compare the hardware it needs.
Choose the model you want to run. Compare hardware that meets its estimated memory needs, not a generic upgrade tier.
Your starting point: 4 GB model budget / 63.8 GB system RAMMemory estimates use the selected quantization and the planner's reserved-memory budget. Context size, runtime support and other applications change actual requirements. This is a shortlist for investigation, not a measured benchmark on the target model or a best-value award.
A desktop-style GPU name does not prove this computer is upgradeable. Check chassis clearance, power supply and connectors, motherboard support, and support for the chosen model on your operating system and inference runtime. Catalog prices can be stale.
The report above gives you the answer first. Open this area only when you want to compare individual models or inspect the repeated measurements behind the score.
Run the same benchmark and get a report grounded in your own hardware, workload conditions, and repeatable local inference.
