Gemma 4 E2B Instruct QAT (UD-Q4_K_XL)¶

Model card

As an immersion guard, blocks 75 in 100 general attacks and 77 in 100 world attacks at a 5% false-block budget. Writes 259 tokens per second and needs 2.8 GB of VRAM at 8K context.

Gemma · E2B · UD-Q4_K_XL · 2.6 GB file · unsloth/gemma-4-E2B-it-qat-GGUF

_{Measured 2026-10-08 · Tryll 0.8.0-DEV · llama.cpp b10839 · NVIDIA GeForce RTX 5070 Ti · vulkan · one deterministic pass (classifiers do not sample) · speed: median of 10 runs · ranked against 13 models}

Plain words¶

What the measured numbers mean for a game. Intervals and ranks are under In the field.

Classification · General guard

75 in 100 · Classification · General guard. Blocks 75 in 100 attempts to pull the NPC out of the game: asking whether it is an AI, jailbreaks, prompt leaks, coding requests and format tricks. Normal player lines wrongly blocked: 8 in 100 on held-out lines (the budget is 5 in 100, set on a separate tuning set).

Classification · World guard

77 in 100 · Classification · World guard. Blocks 77 in 100 lines that do not belong in the setting, such as paying by phone in a medieval town or asking how long chicken keeps in the fridge. Normal player lines wrongly blocked: 1 in 100 on held-out lines (the budget is 5 in 100, set on a separate tuning set).

World guard is measured on one setting, a medieval merchant. A different world needs its own check.

8,200 tokens/s · Performance · Prefill. Reads a 3,889-token prompt in 0.47 s. The first token reaches the player after 0.50 s. A 1,124-token prompt: first token after 0.16 s.

259 tokens/s · Performance · Generation. Writes 259 tokens per second after a 1K-token prompt, and 243 after a 4K-token prompt. 200 tokens take 0.8 s.

2.8 GB VRAM · Performance · Memory. Weights 2.60 GB, KV cache 0.08 GB for 8,192 tokens (q8_0), compute buffers 0.13 GB. Each extra agent on this model adds its own KV cache and compute buffers; the weights are shared.

Speed and memory were measured on one GPU, named in the header. Other hardware gives other numbers.

Not measured yet¶

NPC roleplay: clean speech, staying in lore and role, writing quality
Tool calling: call accuracy and restraint
Summarization: how long the summary keeps what the player said
Intent classification: the butler set is too small for a card number

In the field¶

Where this model sits among the measured models. Right is better. The blue bar is the 95% interval. Guard ranks come from paired comparisons; a speed rank is above another only when the two intervals do not overlap. Close models share a range.

General guard · #2–4 of 13 · setting-invariant classifier

General guard