Model registry

Every model we've calibrated, and how far we trust it.

PRIOR ships with calibration for twenty-two open-weight builds across nine architecture families, up to 14B. This is the whole roster, with each build's steering status stated plainly. Bring your own model and calibrate it in the console; nothing here is a ceiling.

Certified

Cleared for full steering. PRIOR routes green / yellow / red and holds the model to your policy, including forced refusals on RED.

Beta

Runs and gates today: the tri-state detector works, so requests still route by zone. Steering certification is in progress, so load these for the gate, not for guaranteed forced refusals.

Llama

Llama 3.1 / 3.2
Build (GGUF Q4_K_M)ParamsStatus
Llama-3.2-3B-Instruct3BCertified
Llama-3.2-1B-Instruct1BCertified
Meta-Llama-3.1-8B-Instructreference model8BCertified

Qwen3, dense

includes abliterated
Build (GGUF Q4_K_M)ParamsStatus
Qwen3-8B8BCertified
Qwen3-4B-Instruct-25074BCertified
Qwen3-4B4BCertified
Qwen3-14B14BBeta
Qwen3-4B-Instruct-2507abliterated4BBeta

Qwen3-Next

hybrid linear-attention (Gated-DeltaNet)
Build (GGUF Q4_K_M)ParamsStatus
Ornith-1.0-9Bcommunity fine-tune9BCertified
Qwen3.5-9B9BCertified
Qwythos-9Bcommunity fine-tune9BCertified

Phi

Microsoft Phi-4
Build (GGUF Q4_K_M)ParamsStatus
Phi-414BCertified
Phi-4-mini-instruct3.8BCertified

Mistral-Nemo

12B class
Build (GGUF Q4_K_M)ParamsStatus
Mistral-Nemo-12B-Instruct12BCertified
Mistral-Nemo-Instruct-240712BCertified

Ministral

Ministral 3
Build (GGUF Q4_K_M)ParamsStatus
Ministral-3-3B-Instruct-25123BCertified

DeepSeek-R1 Distill

R1 reasoning distill on Qwen-14B
Build (GGUF Q4_K_M)ParamsStatus
DeepSeek-R1-Distill-Qwen-14B14BCertified

Gemma 3 / 4

gate-ready, steering in progress
Build (GGUF Q4_K_M)ParamsStatus
Gemma-4-E4B-it4BeffectiveBeta
Gemma-3-4B-it4BBeta
Gemma-3-12B-it12BCertified
Gemma-4-12B-it12BCertified

SmolLM3

SmolLM3 3B
Build (GGUF Q4_K_M)ParamsStatus
SmolLM3-3B3BBeta

All builds are GGUF Q4_K_M. Steering artifacts (the per-model calibration internals) are part of the licensed package and aren't published here. Grading is ongoing as the fleet grows. Your policy packs are authored model-independently, so they carry to any build unchanged, certified or beta.