Field noteNov 19, 20265 min read1 source

Can I run AI locally? Best models for your GPU

In one minute

CanIRun.ai is a browser-based guide to matching local AI models with a computer’s hardware. It combines device selection, model recommendations, estimated memory use and speed, and launch options—helping builders narrow choices before downloading weights, while leaving actual performance to be tested on their own setup.

What happened

CanIRun.ai presents itself as a tool for answering whether a user can run AI locally. Its page has sections for models, devices, a playground, comparisons, and a tier list. Users can select or detect hardware and explore recommendations across text and coding, image, video, and small-model tasks. The page says 40 of 146 models fit comfortably for the selected AMD Radeon 780M Graphics configuration, 12 are a tight fit, and 94 are too heavy. It also lists hardware options ranging from GPUs and Apple Silicon Macs to mobile chips and Raspberry Pi devices.

For each model, the site displays details such as parameter count, estimated memory use, context length, and approximate speed. The selected Radeon 780M view, for example, shows Qwen 3 8B at 4.6 GB and about 13 tokens per second, and GPT-OSS 20B at 11.3 GB and about 30 tokens per second. Recommendations vary by task: the page separates coding and chat from image and video generation, and offers model filters, license filters, and sorting by measures including score, parameter count, context, speed, and estimated VRAM.

The site also connects recommendations to ways of running models. It references runai, Ollama, and LM Studio, and presents quantization choices for individual model entries. Its model cards identify publishers and licenses—for example, Apache 2.0 for the listed Qwen 3 8B and GPT-OSS 20B entries. The site is therefore not just a catalog of model names: it attempts to translate hardware constraints into a shortlist and a possible local-launch path.

The page includes a large catalog of models and devices, including open-weight text, coding, image, and video systems. The displayed entries provide estimates and descriptors rather than a standardized measured comparison across a stated test suite. For builders, the practical function is discovery and initial fit assessment: select a machine, identify plausible candidates, then verify them in the runtime and workload that matter.

What the archive adds

New Runtime’s earlier coverage described local inference as a question of usable fit, not a binary hardware threshold, including the importance of configuration. CanIRun.ai makes that problem more navigable by putting hardware selection and model recommendations in one interface. This is a new tool on a familiar local-inference thread, not evidence that local model performance has changed.

Trend signal

This connects loosely to the pattern that generative AI becomes more useful when it is composed into applications rather than treated as a one-off model interaction. CanIRun.ai turns a model catalog into a decision interface by filtering recommendations around a device and task. That is a modest product-design signal, not evidence of a broader shift in local inference capability. Generative media becomes application infrastructure is the nearest pattern, though the site’s coverage is broader than media.

Why it matters

Local inference decisions are usually constrained by several factors at once: memory capacity, model size, quantization, runtime behavior, task type, and acceptable speed. A catalog that puts those factors near one another can save builders time during initial exploration and make hardware trade-offs easier to explain. It can also help a team avoid treating a model’s parameter count as a sufficient proxy for whether it will run well.

The tool’s strongest role is as a shortlist generator. A recommendation does not establish that a model will meet latency, quality, or reliability requirements under production load. Teams should treat the displayed fit and speed as hypotheses, then benchmark representative prompts and concurrency on the target machine. That distinction matters especially when comparing local use with hosted APIs: privacy and offline operation may be attractive, but the operational cost includes setup, maintenance, and hardware capacity.

What happens next

  • Builders will use hardware-aware catalogs primarily to narrow a local-model test set rather than to make a final deployment choice. (3 months, 70%). Watch for: recommendations paired with runtime-specific measurements and reproducible configurations.

Limits and uncertainty

The estimates shown on the page are not a substitute for a controlled benchmark on a builder’s own hardware. The displayed example gives approximate memory and speed figures, but does not establish that those figures will hold across runtimes, quantization formats, operating systems, thermal conditions, or prompt lengths. Nor does a fit label establish output quality for a particular task. The page is a discovery and recommendation interface; teams should verify licensing and model details at the publisher’s source before deployment, and measure latency and memory under their own workload.

Skill to build now

Use the site to create a small local-model shortlist for one concrete task: select your actual device, identify two or three candidates, and record the displayed memory and speed estimates alongside the model license. Then run the same representative prompts in your chosen runtime and measure latency, peak memory, and output quality. Keep the recommendation and the observed result separate; that gives your team a repeatable way to decide whether local inference is viable.

Reader takeaway

CanIRun.ai makes the first step in local model selection easier: map a device to plausible models and launch paths. Treat its ratings as a starting point, not a performance guarantee. The useful decision comes after testing the shortlisted models on the hardware, runtime, and task your team will actually use.

Evidence / sources

  1. 1canirun.ai
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicModelsExplore the models topic hub.
  2. 02archiveField NotesOpen the latest editorial analysis.
  3. 03sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract