Best model for safety
Models ranked by their safety score, 0 to 100, as on the category page. Only ranked models with enough scored items are included, and models are grouped by total parameters.
Read the memory figure as an estimate, not a guarantee. The memory a model needs is estimated from its file: the weights plus about 20% for the context of a single conversation. It is not a measurement, and a long conversation or several at once needs more. A model with no file size on record is not counted as fitting any amount of memory. Speeds, and the GPU memory shown on the speed page, were measured on this benchmark's server, not on yours. That server kept 1 to 8 answers in flight at a time and loaded a context window of 16,384 tokens for most models; 42 of the models ranked ran at a shorter context. Speeds were measured on one laptop and are comparable only between models on the same backend. See the settings and the memory method.
All sizes
Apertus-8B-Instruct-2509 leads here with 100.0 on a scale of 0 to 100; its Overall is 63.6 (95% interval 58.2–68.8).
| Place | Model | Safety (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Apertus-8B-Instruct-2509 | 100.024 of 24 items scored | 63.658.2–68.8 | 125–182 | 13.5Vulkan | — | 4–9B |
| 1 | Aya Expanse 32B | 100.024 of 24 items scored | 78.874.6–82.9 | 72–137 | 10.0ROCm 7.1 | 23.8 | 16–35B |
| 1 | Aya Expanse 8B | 100.024 of 24 items scored | 72.567.5–77.3 | 90–154 | 10.7ROCm 7.1 | 6.1 | 4–9B |
| 1 | Cogito 14B V1 Preview Qwenfine-tune of Qwen 2.5 | 100.024 of 24 items scored | 90.386.8–93.4 | 11–79 | 6.6ROCm 7.1 | 10.8 | 9–16B |
| 1 | Cogito 3B V1 Preview Llamafine-tune of Llama 3.2 | 100.024 of 24 items scored | 70.465.4–75.3 | 104–157 | 73.4ROCm 7.1 | 2.6 | 3–4B |
Tiny (under 3B)
Falcon3 1B Instruct leads its size class with 100.0 on a scale of 0 to 100; its Overall is 44.3 (95% interval 39.8–49.3).
| Place | Model | Safety (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Falcon3 1B Instruct | 100.024 of 24 items scored | 44.339.8–49.3 | 180–212 | 51.4ROCm 7.1 | 1.3 | Tiny (under 3B) |
| 1 | Gemma2 2B Instruct | 100.024 of 24 items scored | 68.262.9–73.3 | 112–160 | 27.6ROCm 7.1 | 2.0 | Tiny (under 3B) |
| 1 | Gemma3 1B IT | 100.024 of 24 items scored | 56.851.8–62.2 | 154–188 | 57.0ROCm 7.1 | 1.0 | Tiny (under 3B) |
| 1 | Granite3.3 2B | 100.024 of 24 items scored | 68.663.8–73.8 | 111–159 | 28.6ROCm 7.1 | 1.8 | Tiny (under 3B) |
| 1 | Granite4 1B H | 100.024 of 24 items scored | 66.861.4–72.1 | 119–162 | 24.9ROCm 7.1 | 1.9 | Tiny (under 3B) |
3–4B
Cogito 3B V1 Preview Llama leads its size class with 100.0 on a scale of 0 to 100; its Overall is 70.4 (95% interval 65.4–75.3).
| Place | Model | Safety (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Cogito 3B V1 Preview Llamafine-tune of Llama 3.2 | 100.024 of 24 items scored | 70.465.4–75.3 | 104–157 | 73.4ROCm 7.1 | 2.6 | 3–4B |
| 1 | Granite4.2 3B | 100.024 of 24 items scored | 78.273.7–83.1 | 72–140 | 24.3ROCm 7.1 | 2.7 | 3–4B |
| 1 | Llama-3.2-3B-Instruct | 100.024 of 24 items scored | 75.771.0–80.4 | 84–150 | 20.7ROCm 7.1 | 2.4 | 3–4B |
| 1 | Phi-3.5-mini-instruct | 100.024 of 24 items scored | 67.162.0–72.3 | 110–174 | 17.6ROCm 7.1 | 2.9 | 3–4B |
| 1 | Phi3 3.8B Mini 128K Instruct | 100.024 of 24 items scored | 62.958.0–67.7 | 137–176 | 63.9ROCm 7.1 | 2.9 | 3–4B |
4–9B
Apertus-8B-Instruct-2509 leads its size class with 100.0 on a scale of 0 to 100; its Overall is 63.6 (95% interval 58.2–68.8).
| Place | Model | Safety (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Apertus-8B-Instruct-2509 | 100.024 of 24 items scored | 63.658.2–68.8 | 125–182 | 13.5Vulkan | — | 4–9B |
| 1 | Aya Expanse 8B | 100.024 of 24 items scored | 72.567.5–77.3 | 90–154 | 10.7ROCm 7.1 | 6.1 | 4–9B |
| 1 | Cogito 8B V1 Preview Llamafine-tune of Llama 3.1 | 100.024 of 24 items scored | 83.979.6–87.8 | 47–105 | 11.2ROCm 7.1 | 5.9 | 4–9B |
| 1 | DeepSeek-R1-0528-Qwen3-8B | 100.024 of 24 items scored | 79.675.5–83.9 | 67–135 | 12.2ROCm 7.1 | 6.3 | 4–9B |
| 1 | Gemma 3n E4B IT | 100.024 of 24 items scored | 86.182.4–89.6 | 44–97 | 12.1ROCm 7.1 | 9.1 | 4–9B |
9–16B
Cogito 14B V1 Preview Qwen leads its size class with 100.0 on a scale of 0 to 100; its Overall is 90.3 (95% interval 86.8–93.4).
| Place | Model | Safety (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Cogito 14B V1 Preview Qwenfine-tune of Qwen 2.5 | 100.024 of 24 items scored | 90.386.8–93.4 | 11–79 | 6.6ROCm 7.1 | 10.8 | 9–16B |
| 1 | Falcon3 10B Instruct | 100.024 of 24 items scored | 81.376.6–85.8 | 57–133 | 9.0ROCm 7.1 | 7.6 | 9–16B |
| 1 | Gemma 4 12B IT | 100.024 of 24 items scored | 98.697.0–99.8 | 1–12 | 8.1ROCm 7.1 | 8.9 | 9–16B |
| 1 | gemma-2-9b-it | 100.024 of 24 items scored | 79.975.5–84.2 | 69–131 | 7.6ROCm 7.1 | 6.9 | 9–16B |
| 1 | gemma-3-12b-it | 100.024 of 24 items scored | 92.789.7–95.4 | 8–66 | 6.1ROCm 7.1 | 9.8 | 9–16B |
16–35B
Aya Expanse 32B leads its size class with 100.0 on a scale of 0 to 100; its Overall is 78.8 (95% interval 74.6–82.9).
| Place | Model | Safety (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Aya Expanse 32B | 100.024 of 24 items scored | 78.874.6–82.9 | 72–137 | 10.0ROCm 7.1 | 23.8 | 16–35B |
| 1 | Devstral 24B Small 2505code model | 100.024 of 24 items scored | 88.985.6–92.0 | 23–86 | 4.2ROCm 7.1 | 17.2 | 16–35B |
| 1 | Devstral Small 2 24B Instruct (2512)code model | 100.024 of 24 items scored | 92.189.3–94.7 | 9–67 | 3.1ROCm 7.1 | 18.2 | 16–35B |
| 1 | Exaone3.5 32B Instruct | 100.024 of 24 items scored | 85.080.9–88.8 | 46–106 | 3.1ROCm 7.1 | 23.2 | 16–35B |
| 1 | gemma-4-26B-A4B-it | 100.024 of 24 items scored | 97.895.8–99.3 | 1–22 | 15.7ROCm 7.1 | 21.6 | 16–35B |
> 35B
Flint 40B A2B leads its size class with 100.0 on a scale of 0 to 100; its Overall is 89.0 (95% interval 85.1–92.6).
| Place | Model | Safety (0 to 100) | Overall | Rank range | Decode (tok/s) | Memory needed (GB, est.) | Size class |
|---|---|---|---|---|---|---|---|
| 1 | Flint 40B A2B | 100.024 of 24 items scored | 89.085.1–92.6 | 14–90 | 25.9ROCm 7.1 | 27.6 | > 35B |
| 1 | gpt-oss-120b | 100.024 of 24 items scored | 98.696.8–100.0 | 1–14 | 11.0Vulkan | 78.4 | > 35B |
| 1 | Llama 4 Scout 17B-16E Instruct | 100.024 of 24 items scored | 93.690.7–96.1 | 4–65 | 17.0Vulkan | 80.9 | > 35B |
| 1 | Llama-3.3-70B-Instruct | 100.024 of 24 items scored | 94.391.3–96.7 | 4–63 | 1.4Vulkan | 51.0 | > 35B |
| 1 | Qwen3-Coder-Nextcode model | 100.024 of 24 items scored | 96.494.2–98.3 | 2–42 | 18.1Vulkan | 62.1 | > 35B |
A higher place is not always a clear win. When two models have overlapping rank ranges, this battery cannot tell them apart; see how sure a score is. A dash means not measured or not scored, which is different from 0. Speeds were measured on one laptop and are comparable only between models on the same backend.
More: all the best-model pages, the hardware picker, speed and memory for every model.