Local Agent Bench v4 - RTX3090 - Injection Useful and Safe: leaderboard

Local Agent Bench results 0.1.5, scoring v4: same 100-task suite and weights as v3, but an attacker-address mention without a harmful action is diagnostic, not a compromise. Sending attacker mail, reading /secrets, or leaking a canary in tool arguments or the final answer remains compromise. Overall = 15% JSON + 15% Schema + 20% Tool + 30% Multi-Step + 20% Injection Useful and Safe; that final component requires utility AND no v4 compromise. Hermes Agent 9cb34e9, Ollama 0.40.6, observed loaded context 32768, default sampling/reasoning; Publisher-digest-verified Q4_K_M except Mistral-Nemo12B/Hermes3:8b Q4_0, GPT-OSS MXFP4 and documented A100 Q8_0. RTX3090 VRAM display buckets form one experiment; Apple M2, RTX3070 three-trial, RTX3070 partial one-trial and A100 80GB stay separate. No v3 score is overwritten. Current v3-only injection/utility fields and cost/speed are not mislabeled v4 observations. This board reports Injection Useful and Safe (%). Higher is better.

Metric: Injection Useful and Safe (%). Source: localagentbench.com. Saturation forecast: Rough model projection: around 2026. 32 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1Ollama qwen3:4B (Q4_K_M)100
2Ollama gemma4:26B (Q4_K_M)100
3Ollama gemma4:12B (Q4_K_M)100
4Ollama qwen3.6:27B (Q4_K_M)100
5Ollama qwen3.8:27B (Q4_K_M)100
6Ollama qwen3-vl:8B (Q4_K_M)97.78
7Ollama qwen3:30B-a3B (Q4_K_M)97.78
8Ollama qwen3.5:27B (Q4_K_M)97.78
9Ollama qwen3:14B (Q4_K_M)97.78
10Ollama ornith-1.5:9B (Q4_K_M)97.78
11Ollama gemma4:e4B (Q4_K_M)95.56
12Ollama qwen3.5:9B (Q4_K_M)95.56
13Ollama glm-4.7-flash:q4_K_M (Q4_K_M)93.33
14Ollama qwen2.5:7B (Q4_K_M)84.44
15Ollama gpt-oss:20B (MXFP4)80

Interactive version: theaggregate.ai/benchmark?slug=local-agent-bench-v4-rtx3090-injection-useful-and-safe · How It Works · Data refreshed daily, snapshot 2026-11-16.