Preview

TQ Arena

A side-by-side compare UI for local engines. Type one prompt, send it to several engines or GPUs at once, and watch each answer stream with its own live tokens per second, time to first token and totals. HTML answers open next to each other when they finish. TQ Arena is not released yet; these clips are recordings of the development build.

RTX 3090 vs RTX 5060 Ti, same prompt. Left: RTX 3090 24 GB, Qwen3.8-27B-TQ3_4S-v2.gguf (14.5 GB), 278.9 tok/s. Right: RTX 5060 Ti 16 GB, a 12.55 GB pre-release TQ file (shown as TQ3_M), 150.7 tok/s. Both tqLLM with the DFlash2 drafter, the LocalMaxxing reasoning prompt, 320 output tokens, thinking on (2026-10-06).

RTX 3090 vs DGX Spark, a coding prompt. "Write a complete Python class for a binary search tree". Left: RTX 3090, Qwen3.8-27B-TQ3_4S-v2.gguf, DFlash2 Q4_K_M drafter, 291 tok/s. Right: GB10, an earlier TQ4 file, DFlash2 Q4_K_M drafter, 93 tok/s. Thinking off, temperature 0 (2026-09-30 build). Ends with the saved session restored from History.

HTML side by side. An animated analog clock as one HTML file. Left: RTX 3090, 237.9 tok/s, 1,062 tokens in 4.6 s. Right: GB10, 78.4 tok/s, 1,237 tokens in 16.3 s. Both an earlier Qwen3.8-27B TQ4 file with the DFlash2 Q4_K_M drafter, thinking off (2026-09-30 build). Both pages open running when the answers finish.

A copy task on one RTX 3090. The prompt contains a finished HTML scene and asks the model to return it, so the drafter's guesses are almost always right: 350-365 tok/s as it streams, with Qwen3.8-27B-TQ3_4S-v2.gguf and the DFlash2 drafter (2026-10-04). Edits of existing text behave like this; new writing is slower (see the other clips).

The speeds in the clips are the engines' own timings as shown in the UI. For controlled measurements see Performance.