DeepSeek v4 Flash 0731 for one DGX Spark has been updated, I think many will like this one ✨ It now offers an abliterated (uncensored) path, with no weights change. Just a simple flag. - default stays stock - same checkpoint - same kv cache - same decode tok/s As far as I know this is the only abliterated DS4F repo that fits one DGX Spark with excellent quality. Thanks to @u1tra_instinct for the refusal direction! Get it here:

DeepSeek V4 Flash 0731 is impressive. It feels like a capabilities leap for a model in this size class. It is also incredibly cost efficient. The model is available in Bionic, both for running locally and through the LM Studio Cloud!

DeepSeek-V4-Flash-0731 (all 284B parameters) running on a single DGX Spark. ⚡ ~17 tok/s generation, 40 tok/s prefill 📦 One 80GB GGUF file, mixed IQ2_XXS/Q8 imatrix quant 🧠 256-expert MoE, 6 routed per token 🆓 MIT licensed frontier model (kinda) on your desk 😎 🤗

So we now have 3 flash models DSV4-flash-0731 Qwen3.8-flash-next GLM-5.3-flash I think this is solid, every model should have a faster flash version, this is exactly what i asked for a while ago. So between these three, what do you say is the best when running locally?

DeepSeek-v4-Pro (0813) vs DeepSeek-v4-Flash (0731) both one-shot via Hermes first time that ds4-flash disappoints me. i'm guessing it's a false flag because it's nowhere near its usual quality. ds4-Pro clearly wins here, it's better on all aspects

We independently verified and reproduced deepseek-v4-flash-0731 eval on terminal bench with latest release of ante. The accuracy is almost the same 82.7% with their unpublished harness! It is currently sitting on top of our leaderboard after adjusting grok 4.5 reward hacking.