Local decision models (System One)

Safety classifiers, rerankers, judges and zero-shot classifiers that return a label or score, rated for precision, speed and clarity.

Step 1

Tell us your computer first

Then you will see how many of these 63 decision models run on it, and can filter the list by how well each one runs.

Detect my hardware Pick my GPU or Mac

System One models answer with a label, a score or a probability that code can use directly, instead of writing text. They are the fast, focused judgments behind safety filters, search ranking, answer grading and routing. Chat LLMs do the opposite job: slower, open-ended reasoning and writing.

Precision
How often it decides correctly, from the benchmark on each model page (F1, nDCG@10, RewardBench). Shown as a 0 to 100 tier.
Speed
Time for one decision on your hardware: reading the input, plus any words it writes. Encoders and rerankers take milliseconds; judges that write feedback take seconds.
Clarity
How easy the answer is to use in code. A probability or score is best, then labeled spans, fixed labels, a reasoned verdict, and written feedback before a score.
Usage
The task, what it checks, how much text it reads, its languages and license.

Prefer not to host one? TypeSafe’s Jev is a hosted System One model that returns typed answers and probabilities through an API; everything listed here runs on your own computer.

Model Task Returns Size Precision Speed on RTX 4090 License Runs in
Qwen3 Reranker 4BAlibaba Rerank Probability or score 4.0B 90 52 ms apache-2.0
Qwen3 Reranker 8BAlibaba Rerank Probability or score 8.2B 90 88 ms apache-2.0
Skywork Reward V2 Llama 3.1 8BSkywork Judge Probability or score 7.5B 90 144 ms llama3.1
zerank-2ZeroEntropy Rerank Probability or score 4.0B 88 52 ms apache-2.0
CompassJudger-2 32BOpenCompass Judge Written feedback, then a score 33B 86 4.5 s apache-2.0 +6
Jina Reranker v3.5Jina AI Rerank Probability or score 600M 86 11 ms cc-by-nc-4.0
Qwen3Guard Gen 8BAlibaba Safety Fixed labels 8.2B 86 132 ms apache-2.0 +6
Shieldstral 1.0 3BMistral AI Safety Probability or score 3.9B 86 50 ms apache-2.0
Granite Guardian 4.1 8BIBM Safety Fixed labels 8.4B 84 135 ms apache-2.0 +6
Jina Reranker v3Jina AI Rerank Probability or score 600M 84 11 ms cc-by-nc-4.0
Llama 3.1 Nemotron Safety Guard 8B v3NVIDIA Safety Fixed labels 8.0B 84 129 ms other +6
Qwen3Guard Gen 4BAlibaba Safety Fixed labels 4.4B 84 84 ms apache-2.0 +6
Skywork Reward V2 Qwen3 8BSkywork Judge Probability or score 7.6B 84 146 ms apache-2.0
mxbai-rerank-large v2Mixedbread Rerank Probability or score 1.5B 82 23 ms apache-2.0
GLiNER2.5 DecideFastino Classify Probability or score 490M 80 6.6 ms apache-2.0
gpt-oss-safeguard 120BOpenAI Safety Reasoned verdict 117B (5.1B active) 80 — apache-2.0 +7
Granite Guardian 3.3 8BIBM Safety Fixed labels 8.2B 80 131 ms apache-2.0 +6
Qwen3 Reranker 0.6BAlibaba Rerank Probability or score 600M 80 11 ms apache-2.0
Selene 1 Mini 8BAtla Judge Written feedback, then a score 8.0B 80 1.3 s apache-2.0 +6
WildGuard 7BAi2 Safety Fixed labels 7.3B 80 117 ms apache-2.0
zerank-1 smallZeroEntropy Rerank Probability or score 1.7B 80 25 ms apache-2.0
DynaGuard 8BUniversity of Maryland Safety Reasoned verdict 8.2B 79 1.1 s apache-2.0 +6
BGE Reranker v2 GemmaBAAI Rerank Probability or score 2.5B 78 34 ms apache-2.0
CompassJudger-2 7BOpenCompass Judge Written feedback, then a score 7.6B 78 1.3 s apache-2.0 +6
Nemotron Content Safety Reasoning 4BNVIDIA Safety Reasoned verdict 4.3B 78 842 ms other +6
OpenAI Privacy FilterOpenAI Classify Labeled spans or fields 1.4B (50M active) 78 4.3 ms apache-2.0
gpt-oss-safeguard 20BOpenAI Safety Reasoned verdict 22B (3.6B active) 76 905 ms apache-2.0 +7
Llama Prompt Guard 2 86MMeta Safety Probability or score 280M 76 7.0 ms llama4
mxbai-rerank-base v2Mixedbread Rerank Probability or score 490M 76 9.9 ms apache-2.0
NuExtract 2.0 4BNuMind Classify Labeled spans or fields 3.8B 76 108 ms mit +6
GLiClass ModernBERT Large v3Knowledgator Classify Probability or score 400M 74 6.2 ms apache-2.0
Prometheus 2 8x7BPrometheus Eval Judge Written feedback, then a score 47B (13B active) 74 13.6 s apache-2.0 +6
Qwen3Guard Gen 0.6BAlibaba Safety Fixed labels 750M 74 26 ms apache-2.0 +7
BGE Reranker v2 M3BAAI Rerank Probability or score 570M 72 10 ms apache-2.0 +2
DeBERTa v3 Large Zero-shot v2.0Moritz Laurer Classify Probability or score 440M 72 6.4 ms mit
DuoGuard 0.5BDuoGuard Safety Probability or score 490M 72 9.3 ms apache-2.0
Skywork Reward V2 Qwen3 0.6BSkywork Judge Probability or score 600M 72 17 ms apache-2.0
DynaGuard 1.7BUniversity of Maryland Safety Reasoned verdict 2.0B 70 284 ms apache-2.0 +6
GTE Reranker ModernBERT BaseAlibaba Rerank Probability or score 150M 70 5.6 ms apache-2.0
Jina Reranker v2 Base MultilingualJina AI Rerank Probability or score 280M 70 7.0 ms cc-by-nc-4.0 +2
ModernBERT Large Zero-shot v2.0Moritz Laurer Classify Probability or score 400M 70 6.2 ms apache-2.0
NuExtract 2.0 2BNuMind Classify Labeled spans or fields 2.2B 70 98 ms mit +6
GLiGuard 300MFastino Safety Probability or score 210M 68 6.3 ms apache-2.0
GLiNER2 Base v1Fastino Classify Labeled spans or fields 210M 68 5.1 ms apache-2.0
Granite Embedding Reranker English R2IBM Rerank Probability or score 150M 68 5.6 ms apache-2.0
Llama Guard 3 8BMeta Safety Fixed labels 8.0B 68 129 ms llama3.1 +6
Llama Prompt Guard 2 22MMeta Safety Probability or score 71M 68 4.8 ms llama4
Prometheus 2 7BPrometheus Eval Judge Written feedback, then a score 7.2B 68 1.2 s apache-2.0 +7
GLiClass Base v3Knowledgator Classify Probability or score 190M 66 5.0 ms apache-2.0
GLIDERPatronus AI Judge Written feedback, then a score 3.8B 66 737 ms cc-by-nc-4.0 +6
LettuceDetect LargeKR Labs Judge Labeled spans or fields 400M 66 13 ms mit
Llama Guard 4 12BMeta Safety Fixed labels 12B 66 191 ms llama4
ProtectAI DeBERTa v3 Prompt Injection v2Protect AI Safety Probability or score 180M 66 5.9 ms apache-2.0
ShieldGemma 27BGoogle Safety Probability or score 27B 66 260 ms gemma +6
DeBERTa v3 Base Zero-shot v2.0Moritz Laurer Classify Probability or score 180M 64 5.0 ms mit
Flow Judge v0.1Flow AI Judge Written feedback, then a score 3.8B 64 737 ms apache-2.0 +6
ShieldGemma 9BGoogle Safety Probability or score 9.2B 64 98 ms gemma +6
GLiNER Multi v2.1Urchade Zaratiana Classify Labeled spans or fields 290M 62 5.6 ms apache-2.0
HHEM 2.1 OpenVectara Judge Probability or score 110M 62 6.4 ms apache-2.0
ShieldGemma 2BGoogle Safety Probability or score 2.6B 58 35 ms gemma +6
BART Large MNLIMeta Classify Probability or score 410M 56 6.2 ms mit
Llama Guard 3 1BMeta Safety Fixed labels 1.5B 56 48 ms llama3.2 +6
ms-marco MiniLM L6 v2Sentence Transformers Rerank Probability or score 23M 52 4.2 ms apache-2.0

Speed is the time for one decision on an RTX 4090 at the best-fitting precision, from our estimates. Open a model for any GPU or Mac.

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.