The best uncensored Microsoft model is Phi 4 (4.9 intelligence) — served without a provider-level moderation layer.
Large language models served without a provider-level moderation layer, ranked by intelligence. "Uncensored" here means the API endpoint doesn't add its own filtering on top of the model — each model's own alignment training still applies, and open-weight entries can be self-hosted for full control. Relevant for fiction, roleplay, red-teaming and safety research, where an extra moderation layer causes over-refusal on legitimate prompts.
The best uncensored Microsoft model is Phi 4 (4.9 intelligence) — served without a provider-level moderation layer.
modelgrep tracks 2 Microsoft models with live benchmarks, speed, latency and per-provider pricing, led on intelligence by Phi 4. 1 of them qualify for this ranking.