LocalLLMChecker

Update log

Every change that moves a number you might rely on, newest first. That includes our mistakes, because an estimate you cannot check for errors is not worth much.

Engine

Measured validation, and an engine rebuilt around it

We ran Qwen2.5 7B on an RTX 3050 6 GB across five context sizes and six offload levels and compared the results with the calculator. The first version was wrong in three ways: it overestimated memory by 8 to 18 percent, it said a 7B model did not fit a 6 GB card that held it with half a gigabyte to spare, and it mixed binary and decimal gigabytes in the speed formula. All three are fixed. On this machine memory is now off by about 2 percent and speed by about 2.5 percent on average.

The CUDA and CPU efficiency constants are now the measured values, 0.79 and 0.63. The full results, the raw data and the script that produced them are on the validation page.

Engine

Attention is read layer by layer

The KV cache used to assume one attention design per model. It now reads the kind of every layer from the configuration: full, sliding window, linear or compressed. That corrects the cache of hybrid models such as Qwen3.5, which we had overstated by about 4 times, and of models that alternate full and windowed layers, such as gpt-oss and Gemma, which we had understated or overstated depending on the case.

The catalog was cut from 133 to 82 models. We removed research baselines, very old and very small models, and every model whose attention layout could not be established from its configuration, since a memory figure we cannot stand behind is worse than none.

Engine

Mixture-of-experts models: expert offload and a speed correction

When a mixture-of-experts model does not fit in VRAM, the engine now also considers keeping its experts in system RAM and everything dense on the GPU, which is what llama.cpp does with --cpu-moe, and reports whichever is faster.

MoE speeds were too optimistic by about 2 times, because gathering small expert matrices wastes bandwidth that the plain formula counts as used. We now apply a factor of 0.58 to the GPU bandwidth for these models, set from published llama.cpp results. We have not measured a MoE model ourselves and say so wherever those speeds appear.

Catalog

Native low-precision formats use their real file size

gpt-oss-20b and gpt-oss-120b ship their experts in MXFP4, so a download is the same size whatever the file is called. We now use the file size read from the repository, 11.3 GiB and 59 GiB. The 20B model had been shown as fitting a 12 GB card with room to spare, and it does not: its weights alone take nearly all of it.

Engine

Recommendation scoring changed

The "best model for your card" lists now treat a mixture-of-experts model as sitting between its active and total size, favor newer and more widely downloaded models slightly, and offer only general-purpose models as the all-round pick. Code-only, vision-only and very small models still appear in the full tables.

Capability tags that Hugging Face does not provide were filled in by hand for 35 models, and are kept in a curation file so a catalog refresh does not lose them.

Correction

The calculators on the site now use the same data as the static pages

The live calculators on the home, compare and buying pages were sent a trimmed copy of the catalog that left out three fields the engine reads: whether the embedding is tied, and the expert counts of mixture-of-experts models. Their answers could differ from the static page next to them. They now share one packing function, and a test checks that every model gives the same answer before and after packing.

Catalog

Code models added to the catalog

Qwen2.5 Coder 7B, 14B and 32B and Qwen3 Coder 30B A3B are now listed as models in their own right. The Qwen2.5 and Qwen3 Coder models share an architecture with their general-purpose siblings and need the same memory, but they answer a different question, which is what to run for code, so they have their own pages and appear in the coding picks.

Content

Guides section

We added a library of guides on how much VRAM you need, quantization, the KV cache, memory bandwidth, buying hardware, running models in llama.cpp and Ollama, and image, video and speech generation. Each one is built from the same engine as the calculator, so its numbers cannot drift from it.

Catalog

Image, video and audio catalog refreshed

The media catalog was rebuilt from Hugging Face, with component sizes read from the published files.

Catalog

Language model catalog and prices refreshed

The language model catalog was rebuilt from Hugging Face and the price file was reviewed by hand.

The measurements behind the engine are on the validation page, and the formulas are in the methodology. If you find a number that disagrees with your machine, the contact page says what to send.