Models

A page per model worth running locally - grade, decode taper, VRAM, and the serving gotchas that actually decide whether it works. Every number is read live from the benchmark, so these stay current as models are re-measured. For the whole field at once, see the data page.

GLM-5.3-Flash MoE + sparse-MLA· W4A16-NVFP4 experts, FP8 attention

The new daily driver, and the strongest model that fits on the two-card box: on the full benchmark it comes second only to Opus, edging out DeepSeek V4 Flash for the top local spot, though the error bars still overlap so the lead is a narrow one. It is also the heaviest local model, and it only runs at all through a build quantized for this exact hardware.

2.56

mean grade

DeepSeek V4 Flash (0731) sparse-MLA MoE· DSpark spec-decode

The workhorse that held the daily-driver slot until GLM narrowly took it: a sparse-attention MoE that holds its decode speed flat across the whole context window, and, after a July re-post-train, one of the most reliable local models on the benchmark.

2.43

mean grade

Qwen 3.8 27B dense 27B· four-bit NVFP4

A dense 27B that thinks about twice as hard as anything its size. Quantized to four bits it fits in 23 GB and ties DeepSeek V4 Flash on the coding benchmark - and the four-bit build turned out to be the one to serve.

2.42

mean grade

Laguna-S ~67 GB weights· thinking model

The sleeper: level with DeepSeek V4 Flash on the hard tasks at 67 GB of weights - under half the footprint. Matching DeepSeek at this size is the standout of the batch.

2.28

mean grade