GLOBALATLAS AI · DEDICATED INFERENCE

DeepSeek-R1-Distill-Qwen-14B

DeepSeek-R1-Distill-Qwen-14B: habayn inference oo iskiis loo marti-geliyo, weights BF16, context 8192-token ah, xusuus GPU iyo xulashada server-ka.

Qaabeynta

Model reasoning oo gaar ah oo DeepSeek ah, lagana soo miiray wax-soo-saarka DeepSeek-R1 oo ku salaysan asaas Qwen2.5-14B ah. Ma aha model-ka buuxa ee 671B DeepSeek-R1 ama xirmo ka kooban laba model.

Moodel: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
Parameters / weights: 14.77B · — GB
Qaabka weights-ka: BF16
Context / concurrerncy: 8192 / 1
Xadka context-ka checkpoint: 131072
Xusuusta GPU ee la qiyaasay: Requires an updated assessment
RAM / disk la qorsheeyey: ≥ 64 GB / ≥ 150 GB
Runtime serving: vLLM · HTTP API
Shati: MIT; model aasaasi ah Apache-2.0

Qiyaastu waxay ku dartaa weights BF16, KV cache, goob shaqo 2 GB ah iyo kayd xad-dhaaf ah oo 15% GPU ah. RAM iyo disk-ga waa shuruudo qorsheyn GlobalAtlas, ma aha minima rasmiga ah ee horumariyaha. Nooca disk-ga, drivers-ka iyo runtime-ka waa la xaqiijiyaa ka hor rakibidda. Miisaaniyadda serving-ka ee 8192-token waxay ku jirtaa gelinta iyo soo saarista; context dheer ama codsiyo isku mar badan waxay u baahan yihiin cabbir beddelid iyo load test.

Kaarka asalka ah ee model-ka

Dooro kaabayaashaaga

Dooro hardware-ka, ka dibna ku xaqiiji helitaanka iyo rakibidda model-ka khabiir. Dalabka server-ka oo keliya ma rakibo model-ka.

Waxaan kula dooran doonnaa habayn iswaafaqsan model-kan.

Codso dejinta oo dhamaystiran

Ku dar model-ka, gobolka iyo culeyska la filayo. Khubaradayadu waxay xaqiijin doonaan hardware-ka, rakibidda, ilaalinta gelitaanka iyo qiimaha kama dambaysta ah.

Kala hadal Atlas →

Codso caawimo khabiir →

U dir shuruudahaaga khubarada GlobalAtlas si loo sameeyo dib-u-eegis habayn gaar ah. Waxay hubin doonaan hardware-ka, iswaafajinta iyo shuruudaha, ka dibna waxay kaga jawaabi doonaan iimayl.