Qwen3.8-27B Uncensored Tops 110 t/s Locally

The impressive part is not just the headline speed—it is the combination of agent usability, local privacy, and consumer-grade AMD hardware. Still, the result is highly dependent on quantization, runtime, context length, and speculative-decoding configuration.

• –Qwen3.8-27B is a dense, multimodal, Apache 2.0 model with a 262K-token context window; the uncensored derivative reduces refusals through abliteration while retaining the MTP head. [Model card](https://huggingface.co/JonathanColetti/Qwen3.8-27B-Uncensored)
• –The 7900 XTX result is credible in magnitude: a related ROCm benchmark achieved roughly 102 tokens/s at 9K context and 90 tokens/s at 100K using MTP and optimized kernels. [ROCm benchmark](https://www.reddit.com/r/ROCm/comments/1xdu4mu/qwen3827b_at_100_toks_on_a_single_7900_xtx/)
• –The published BF16 checkpoint needs roughly 55GB of VRAM, so a 24GB 7900 XTX necessarily relies on quantization or another memory-saving representation.
• –“Uncensored” should be read as reduced refusals, not guaranteed compliance or absent safety behavior; the model card explicitly reports residual refusals and recommends adding an application-level safety layer.
• –Developers should reproduce the claim with the exact quant, runtime, context length, MTP acceptance rate, and agent task suite before treating 110 tokens/s as a general-purpose baseline.