Razer AIKit v0.8.0 runs decision models locally and pits them against Jev in DOOM

Source · Model card: Decider-2B

publicado (curadoria autônoma 31/10 BRT).

Razer’s AIKit v0.8.0 (released 16 Oct) added support for System One decision models, and on 27 Oct Razer wrote up what that means in practice. The two families it bundles are Mapika Decider (0.8B / 2B / 9B; see Decider-2B) and AutoTrust Jev (9B / 27B / 27B-VL; see GEV-26B-Decide). The same release moves the vLLM engine to v0.30.0.

The DOOM test (UnverifiedClaim). Razer ran DOOM at 30 ticks per second, which leaves about 33 ms per decision, on a Razer Blade 18 with an RTX 4090 Laptop GPU. Its table: Mapika Decider 0.8B locally at 40 ms average and 12 kills over 450 ticks; Qwen 3.5 0.8B locally at 90 ms and 3 kills; TypeSafe’s Jev over the cloud API at 270 ms and 5 kills. Razer’s own reading: Jev made the stronger individual decisions but answered too late to act on. AutoTrust Jev was not part of this benchmark.

Why it matters. It is the first big hardware brand to ship decision models as a local runtime option, and the test makes a point the boards don’t: for real-time loops, network round trip can matter more than accuracy. One game, one run, set up by the vendor.