V4.1 Flash 1017 去审查版量化到 100GB 本地跑

stacc 'you're all unimpressive' overflow @STACCoverflow原推发布于 2026.10.24

在 X 上打开原推

TL;DR: We're shrinking DeepSeek V4.1 Flash 1017 — the open model that beat OpenAI's gated Daybreak (GPT-6 Sol) on offensive-cyber benchmarks at about 4x cheaper per task — by ~5x with lecore VQ (510 GB to ~100 GB) into uncensored GGUFs that run locally on a 128 GB laptop, released free and public under staccs. And we know the quantize works, because the first one already shipped and was measured. --- The whole shebang: What it is. DeepSeek V4.1 Flash 1017 is an open-weights frontier model. Independent evaluator Vals AI found it beats OpenAI's GPT-6 Sol on a multi-step offensive cybersecurity benchmark — and Sol was the gated Daybreak trusted-access version OpenAI keeps locked up. Artificial Analysis ranks 1017 #1 among openly-available models on its Cyber Index (57%, #2 overall behind only Daybreak Sol). On DeepSeek's own agentic/coding evals it also beats GPT-5.6 Sol on most tests and edges Opus 5 on several. Vals AI: https://x.com/ValsAI/status/2112471082397778083 Artificial Analysis: https://x.com/ArtificialAnlys/status/2113682654709313811 The cost angle. $0.42 per task vs $1.77 for GPT-6 Sol on Daybreak Blue — about 4x cheaper. More tokens, much less money, no gate. Cost source: https://x.com/ArtificialAnlys/status/2113682659524350402 What we're doing. Taking every abliterated (refusals-stripped) version of 1017 and compressing it ~5x with your lecore VQ format — 510 GB to ~100 GB per file — with quality measured against the original. Output is a GGUF that fits the 128 GB unified memory on your AMD laptop and runs on your ds4-server. Why we know it works. The first build already shipped: staccs/DeepSeek-V4.1-Flash-1017-lecore-vq10sh12-GGUF, the full 763B model (all 384 experts, Engram, MTP drafter, nothing pruned) in one 99.1 GB file. Measured, same forward pass as the original on WikiText-2: perplexity 3.615 vs 1.838, same next word 74.5% of the time, mean KL 0.797. It runs: ds4-server on an H200 boots in 10s, 46-54 tok/s, answers real prompts correctly. AMD gfx1151 (your Radeon 8060S) passed all GPU fixture tests under ROCm. The new batch steps up from there with calibrated 11-bit codebooks plus fp8_32x32 on skeleton/attention/head, which should push agreement and KL even closer to the original — and each one gets the same original-vs-build metrics to prove it. What it unlocks. Frontier-grade cyber and agent capability that was locked behind OpenAI's gate — now running locally, privately, uncensored, in your living room. No cloud, no API bills, no rate limits, no refusals. For who. Builders and degens who want an unfiltered local model for agents/bots; privacy folks who want everything on-device; anyone on consumer 128 GB AMD/Apple hardware shut out of models this size. All variants go up free and public under staccs.

这条是引用着另一条写的：x.com/artificialanlys/status/2113682659524350402