RAG Gauntlet — rag-gauntlet-v1
Code Generation Tool Calling RAG / Q&A Reasoning / Math | About

The Gauntlet

Six escalating tests, run identically across every model. Find what actually holds up on your hardware — organized by VRAM tier, not marketing size class.

Sub-4GB

6 models tested
gemma4:e2b Champion
1.8 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 132.4 tok · 2.2s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 248.0 tok · 1.56s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 316.0 tok · 1.88s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 5/5 · 481.6 tok · 2.72s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 392.0 tok · 2.28s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 202.0 tok · 1.32s
30/30
3.4 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 227.0 tok · 3.76s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 305.2 tok · 2.54s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 56.4 tok · 0.73s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 5/5 · 381.4 tok · 3.05s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 43.0 tok · 0.63s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 223.0 tok · 1.96s
30/30
2.7 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 135.2 tok · 2.02s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 458.8 tok · 2.01s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 315.0 tok · 1.45s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 5/5 · 3752.4 tok · 15.84s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 306.0 tok · 1.44s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 208.0 tok · 1.03s
30/30
2.9 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 11.0 tok · 0.48s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 16.0 tok · 0.23s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 19.0 tok · 0.18s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 43.0 tok · 0.3s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 37.0 tok · 0.28s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 54.0 tok · 0.38s
25/30
3.1 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 65.0 tok · 0.78s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 47.0 tok · 0.34s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 35.0 tok · 0.29s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 98.0 tok · 0.55s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 37.0 tok · 0.3s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 17.0 tok · 0.24s
25/30
3.7 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 11.0 tok · 1.6s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 48.6 tok · 0.39s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 20.0 tok · 0.26s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 134.4 tok · 0.78s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 39.0 tok · 0.34s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 17.0 tok · 0.26s
25/30

4-8GB

4 models tested
6.3 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 133.0 tok · 1.75s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 198.8 tok · 1.51s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 236.6 tok · 1.77s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 1676.0 tok · 11.96s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 425.8 tok · 3.09s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 222.6 tok · 1.71s
25/30
mistral:7b Champion
5.6 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 16.0 tok · 0.56s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 35.0 tok · 0.25s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 24.0 tok · 0.2s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 47.0 tok · 0.34s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 46.0 tok · 0.33s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 21.0 tok · 0.21s
25/30
5.1 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 13.0 tok · 0.75s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 42.0 tok · 0.4s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 22.0 tok · 0.28s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 24.0 tok · 0.29s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 47.0 tok · 0.45s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 19.0 tok · 0.28s
25/30
5.7 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 373.0 tok · 19.24s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 770.0 tok · 13.68s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 0/5 · 4096.0 tok · 36.87s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 4096.0 tok · 33.21s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 925.0 tok · 7.62s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 457.0 tok · 3.9s
20/30

8-12GB

4 models tested
deepseek-r1:14b Champion
10 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 243.6 tok · 4.04s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 334.0 tok · 4.12s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 258.8 tok · 3.23s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 5/5 · 645.6 tok · 7.91s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 240.0 tok · 3.03s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 261.0 tok · 3.37s
30/30
10 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 184.0 tok · 4.36s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 323.0 tok · 3.85s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 290.0 tok · 3.47s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 1/5 · 3900.2 tok · 46.25s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 241.4 tok · 2.91s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 199.0 tok · 2.48s
26/30
8.0 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 196.0 tok · 3.59s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 242.8 tok · 3.07s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 203.0 tok · 2.61s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 3664.8 tok · 42.6s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 350.2 tok · 4.24s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 184.0 tok · 2.53s
25/30
8.4 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 146.2 tok · 3.25s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 295.0 tok · 3.81s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 199.0 tok · 2.67s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 4096.0 tok · 49.76s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 370.6 tok · 4.66s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 142.0 tok · 2.13s
25/30

12GB+

3 models tested
15 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 14.0 tok · 2.32s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 17.0 tok · 0.48s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 21.0 tok · 0.55s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 5/5 · 228.8 tok · 4.18s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 47.0 tok · 1.0s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 20.0 tok · 0.59s
30/30
22 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 186.0 tok · 6.93s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 288.0 tok · 7.26s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 275.0 tok · 6.86s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 1/5 · 1185.2 tok · 29.17s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 254.0 tok · 6.37s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 255.0 tok · 6.51s
26/30
17 GB measured VRAM
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 1 — floor: 5/5 · 164.8 tok · 3.92s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 2 — restraint (refusal): 5/5 · 334.0 tok · 2.84s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 3 — distractor resistance: 5/5 · 271.8 tok · 2.38s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 4 — multi-passage synthesis: 0/5 · 1593.6 tok · 12.43s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 5 — precise extraction: 5/5 · 469.4 tok · 3.89s
[This tool call rendered an interactive widget in the chat. The user can already see the result — do not repeat it in text or with another visualization tool.] Tier 6 — long-context needle: 5/5 · 238.0 tok · 2.26s
25/30