Matching posts

Matching terms are highlighted

POST

DeepSeek V4.1 Flash on a 128GB Mac: 1.75-bit quant at 50–76 tok/s

…和数学的精度保留很好，Tool-calling 跑了 92/100；长篇小说和发散创作能看出一些量化退化。 说白了，这不是给日常轻量使用的，而是把统一内存与高速 SSD 带宽利用到极致的方案。 手…

deepseek-v4.1@ohayou792　Filed in No.020

1 day agoCopy original

POST

llama.cpp adds DeepSeek V4.1 Flash, but smallest GGUF is 246 GiB

Upstream llama.cpp b11870 supports DeepSeek V4.1 Flash, but the author found the smallest working community GGUF (Q2_K) needs 246.3 GiB, far beyond 128 GB, and warns some fork GGUFs aren't interchangeable.

本地部署@wikiwayne　Filed in No.020

1 day agoCopy original

DeepSeek V4.1 Flash GSQ-RCO GGUF quantization released

Links to ISTA-DASLab's GSQ-RCO GGUF quantized weights of DeepSeek V4.1 Flash 1017 and the accompanying Abyss repository.

本地部署@rS_alonewolf　Filed in No.020

1 day agoCopy original

POST

DeepSeek V4.1 Flash native quant runs at solid speed across 4x DGX Spark

A user pushes back on the idea that only small 4-bit quants run locally, saying the full native quant of DeepSeek V4.1 Flash 1017 runs at solid speeds across 4x DGX Sparks.

deepseek-v4.1@someRandomDev5　Filed in No.020

20 hours agoCopy original

DeepSeek V4.1 Flash at 2.25 bpw hits 47 tok/s on an RTX 5090

Local DeepSeek-V4.1-Flash at ISTA-DAS 2.25 bpw on one RTX 5090: 47 tok/s long generation and 1,400 tok/s prefill, beating Project Maya on the same box.

deepseek-v4.1@lyesbahi　Filed in No.020

19 hours agoCopy original

POST

Project Maya v1.1.30: 428B MiniMax M3.1 Flash quants fit a 128 GB PC

Local deployment: Maya-XS (93 GB) keeps every expert and Maya-S (122 GB) stays closest to the original, running on V100+ NVIDIA, Strix Halo and llama.cpp, plus a new loop guard against repeating answers.

minimax-m3.1@PeasantSmith　Filed in No.021

1 day agoCopy original

MiniMax M3.1 Flash 428B quantized to 143.9GB for single-machine local runs

…v1.1 73.7%，OSWorld V2 61.7，CyberGym 82.2。 MoE加激进量化，让前沿级agent模型第一次进了高配工作站的范围。按token收费的闭源API，护城河又薄了一…

minimax-m3.1@czw88888　Filed in No.021

15 hours agoCopy original

DeepSeek V4.1 Flash 1017 TensorFold quantized weights on Hugging Face

A Hugging Face repo hosts TensorFold-quantized DeepSeek-V4.1-Flash-1017 weights aimed at running locally on M5 Ultra hardware.

deepseek-v4.1@TechMDAI　Filed in No.020

15 hours agoCopy original

POST

Testing MXFP8 Engram weights instead of 4-bit in a DeepSeek V4.1 Flash 1017 IQ2_XXS quant on Strata

The author tries a DeepSeek V4.1 Flash 1017 IQ2_XXS-Uncensored quant on Strata, swapping the SSD-resident Engram part from 4-bit to original MXFP8 weights using only the weights already on hand.

deepseek-v4.1@Tono_Ken3　Filed in No.020

14 hours agoCopy original

Grok Bot builds a quant trading terminal, turning $44 into $11.6k

Grok Bot 做量化有点强啊 ! ! ! 海外博主搭的交易终端丢了 44.82 美元，一觉醒来账户变成 11,643…

personal-agent@0xlangeai　Filed in No.012

1 day agoCopy original

Showing10of60posts

10 per batch

Keep scrolling to reveal posts 11–20

Showing the first 60 matches. Narrow the query with a model name or tag to find more specific results.