Routed-expert requantisation of `deepseek-ai/DeepSeek-V4.1-Flash` revision `2bc89ac599031fa673cab993f1df02fc4a98c673`, produced on a 32-vCPU Runpod CPU pod with the pinned llama.cpp ggml reference quantisers (PR `ggml-org/llama.cpp#28696` head `cd628010bc3fc0a787d156c969d52a0789451c96`). Tensor names follow that PR's `deepseek4` table. This is a **weights artifact**. No runtime has yet been shown to execute it end to end; the "Runtime status" and "Smoke-test findings" sections say exactly what has been measured on which hardware, and what is still blocked. `SHA256SUMS` covers the five shards. The pack is deterministic: two independent runs produced byte-identical shard digests.

번역 대기 중 · 원문 표시

텍스트 생성

입력 텍스트→ 출력 텍스트

요금 미확인멀티모달 및 채널 요금 별도 확인

자세히 보기 ↗

비교 추가출처 / 문서 ↗

DeepSeek-V4-Flash-0731-Abliterated-DS4-Headroom128

apetersson

> **Experimental low-bit model-surgery derivative.** This release is designed > to run DeepSeek-V4-Flash-0731 with its real three-stage DSpark drafter on a > 128 GB Apple Silicon machine. Structural validation has passed, but direct > generation, quality, refusal-removal, DSpark-acceptance, throughput and peak > memory measurements for this exact artifact are still pending. This is a DS4-native mixed-precision GGUF conversion of `apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8`, which is itself a rank-1 abliterated derivative of `deepseek-ai/DeepSeek-V4-Flash-0731`. It remains a derivative of the official post-trained 0731 model: it is not a base model, was not gradient-finetuned, and is not claimed to be universally "uncensored." The two GGUF files total approximately 86.34 GiB. The profile deliberately spends precision on dense paths used for every token while aggressively compressing the much larger routed-expert bank, leaving practical unified- memory headroom for DS4 runtime state and KV cache on a 128 GB host.

번역 대기 중 · 원문 표시

텍스트 생성

입력 텍스트→ 출력 텍스트

요금 미확인멀티모달 및 채널 요금 별도 확인

자세히 보기 ↗

비교 추가출처 / 문서 ↗

DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128

apetersson

基础档案已收录，详细介绍后续补充。可先查看厂商、分类与来源。

기본 정보 등록 · 상세 정보 보완 예정

텍스트 생성

입력 텍스트→ 출력 텍스트

요금 미확인멀티모달 및 채널 요금 별도 확인

자세히 보기 ↗

비교 추가출처 / 문서 ↗

DeepSeek-V4-Flash-0731-Abliterated-FP8

apetersson

基础档案已收录，详细介绍后续补充。可先查看厂商、分类与来源。

기본 정보 등록 · 상세 정보 보완 예정

텍스트 생성

입력 텍스트→ 출력 텍스트

요금 미확인멀티모달 및 채널 요금 별도 확인

자세히 보기 ↗

비교 추가출처 / 문서 ↗

DeepSeek-V4-Flash-0731-Abliterated-MLX-Mixed-2bit-3bit-g64

apetersson

基础档案已收录，详细介绍后续补充。可先查看厂商、分类与来源。

기본 정보 등록 · 상세 정보 보완 예정

텍스트 생성

입력 텍스트→ 출력 텍스트

요금 미확인멀티모달 및 채널 요금 별도 확인

자세히 보기 ↗

비교 추가출처 / 문서 ↗

DeepSeek-V4-Flash-Vision-Exp-Abliterated

apetersson

> [!WARNING] > Experimental research release. Text and focused image smoke tests pass, but > broader quality, safety, and production-runtime evaluation remains pending. This catalog contains the native abliterated checkpoint and two DS4 GGUF profiles. Both GGUF receivers use tensor types accepted by `antirez/ds4` main; neither requires the legacy Python vision sidecar for current native vision. `Basic128-Routed-IQ2_M` uses the closest familiar BPW class for the complete bundle: 2.915 BPW overall, close to llama.cpp's documented 2.9294 BPW for IQ2_M. The routed-expert subset averages 2.715 BPW, close to IQ2_S. This is a custom mixed recipe rather than llama.cpp's exact IQ2_M preset: routed gate/up tensors use IQ2_XXS, routed down tensors use Q2_K, ten routed layers retain MXFP4, and non-routed tensors include F32, F16, and Q8_0. See the llama.cpp quantization table for the reference BPW classes.

번역 대기 중 · 원문 표시

시각 질의응답

입력 텍스트 · 이미지→ 출력 텍스트

요금 미확인멀티모달 및 채널 요금 별도 확인

자세히 보기 ↗

비교 추가출처 / 문서 ↗