Local Builders Race to Squeeze DeepSeek’s 1017 Weights

@sudoingX and @MiaAI_lab both posted new numbers for DeepSeek V4.1 Flash 1017.

@sudoingX rebuilt his DeepSeek V4.1 Flash weights on the 1017 update. Version 1.2 is live on 2x DGX Sparks.

He says the quality held. His 2.70-bit weights scored a KLD of 0.0243 from native 1017 on his own KL panel, 20% closer than LibertAIDAI’s 2.9-bit pack. Top-1 agreement is 95.47%, and HumanEval+, MBPP+ and IFEval are within noise. On IFEval replies, he writes, “theirs wins, 0.1135 vs 0.1199.”

A minor update that changed every shard.

DeepSeek shipped 1017 as a minor update, so @sudoingX diffed it against his old checkpoint. The config, tokenizer and tensor layout match byte for byte, but every shard’s content is new. In the slices he sampled, about 27% of the 4-bit expert codes flipped.

He pulled the 510 gb checkpoint again and measured all 15,360 experts. A 16-hour encode followed, with one Spark down for almost 3 hours. About 25 hours passed from download to serving. The release is 196 gb of weights plus DeepSeek’s own 1017 engram shards.

The new weights let the built-in drafter accept 4.31 tokens a round on his code reply, against 4.41 before. Next, he is training his own drafter.

@MiaAI_lab’s EAQ v1 posts its own numbers.

@MiaAI_lab says her EAQ v1 quant is 11% to 38% closer to the original than LibertAIDAI’s EXL3, and the gap seems bigger on longer context. She calls LibertAIDAI’s quant “a solid quant” and a strong baseline. EAQ v1.1, v1.2 and EAQ v2 are in the works.

She also shipped hotfixes for her 2x DGX Spark recipe. Long prompts no longer freeze chats, prompts with images now hit the prompt cache, and a looser memguard stops brief memory dips from crashing the model.