Matching posts

Matching terms are highlighted

POST

DeepSeek V4.1 Flash local deploy on 4090 D: 46.2 tok/s with 128GB RAM

project-maya on a 48GB 4090 D raised decode from 25.7 to 46.2 tok/s after adding RAM to 128GB; an engine upgrade dropped it to 37.6 until the PCIe expert path was disabled.

deepseek-v4.1@WenhuanLiu1　Filed in No.020

1 day agoCopy original

POST

Deployment: DeepSeek V4.1 Flash fails to start on 4x RTX 6000 Pro in mainline vLLM

A user complains vLLM merges PRs for non-mainstream setups too slowly, citing V4.1 Flash being completely broken and unable to start on 4x RTX 6000 Pro with mainline vLLM.

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

17 hours agoCopy original

DeepSeek V4.1 Flash on 4x RTX 6000: 340 tok/s single stream in NVFP4

A user's revised inference recipe runs DeepSeek V4.1 1017 in NVFP4 on four RTX 6000 GPUs at 340 tok/s single stream, up from 160 tok/s with an incomplete config.

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

19 hours agoCopy original

POST

Local deployment idea: running DeepSeek-V4.1-Flash-1017 on a DGX Spark

Suggests running DeepSeek-V4.1-Flash-1017 locally on a single DGX Spark alongside GLM-5.3 Flash and Qwen3.8-Flash-Next, possibly covering everything locally.

deepseek-v4.1@sounisi5011　Filed in No.020

2 days agoCopy original

POST

MiniMax M3.1 Flash local deployment on Strix Halo: 2-bit build still needs ~100 GiB

A user sizes up M3.1 Flash for Strix Halo: 428B total / 23B active and 235 GiB of weights mean even a 2-bit build is ~100 GiB before context, one class above what 128GB runs well.

minimax-m3.1@JCCrypto007　Filed in No.021

1 day agoCopy original

POST

Project Maya v1.1.29 adds MiniMax M3.1 Flash support with GGUF quantizations

MiniMax-M3.1-Flash 本地部署又有新进展了！ Project Maya v1.1.29 已正式支持，GGUF 量化版本也已经出…

minimax-m3.1@xulijian1994　Filed in No.021

1 day agoCopy original

POST

Hardware bar for local V4.1 Flash: 8x H200 or 4x RTX PRO 6000

… H200或者社区实验方案4× RTX PRO 6000 Blackwell 现在来说，千问才是本地部署的神。 设备还是太贵了，期待能技术进一步优化吧，让真正的家用消费级显卡能部署Deepseek v…

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

3 days agoCopy original

POST

Local deploy notes: DeepSeek V4.1 Flash ISTA IQ2_XS pack runs in Maya unchanged

A user documents copying the deepseek-v4.1-flash-ist-das-2.25bpw pack into Maya with a new config, running across four GPUs with no code or file conversion.

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

2 days agoCopy original

POST

Local deployment take: DeepSeek V4.1 is too heavy, under 50 tok/s

A developer says only Qwen 3.8 Flash Next + Strata is practical on tok/s, while DeepSeek V4.1 is good but too heavy, running under 50 tok/s and needing an RTX PRO 6000 or better.

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

2 days agoCopy original

POST

MiniMax M3.1 Flash 428B on two DGX Sparks: EXL3 2.15bpw, 39 tok/s single stream

SAGE EXL3 2.15bpw quantization with TP2 over RoCE on two DGX Sparks gives 300K context, 726K KV pool, up to 39 tok/s single stream and 65–78 tok/s at four users.

minimax-m3.1@Tech2Wild　Filed in No.021

4 days 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.