不同LLM分析SDC得效果

Tue, 24 Mar 2026 23:34:34 +0800

5256 input token	tokens消耗		效果
QWQ-32B	5K	2.22 tok/sec • 547 tokens • 151.99s to first token • Stop reason: User Stopped	勉强
grok-3-gemma3-12b-distilled	7K	1.91 tok/sec • 296 tokens • 74.96s to first token •Stop reason: User Stopped 5.95 tok/sec • 371 tokens • 102.78s to first token • Stop reason: User Stopped	勉强
claude-3.7-sonnet-reasoning-gemma3-12b.Q8_0	14K	3.42 tok/sec•1096 tokens•326.91s to first token•Stop reason: EOS Token Found	小错误
QWEN3 14B	8K	21.32 tok/sec•3545 tokens•5.63s to first token•Stop reason: EOS Token Found	输出最好的输出速度感觉不错
Lmstudio/deepseek-r1-0528-qwen3-8b	12K	3.56 tok/sec•7456 tokens•51.63s to first token•Stop reason: EOS Token Found	有点答非所问错误很少
google/gemma-3-27b	9K	0.95 tok/sec • 2292 tokens • 179.37s to first token •Stop reason: EOS Token Found	错误太多，简直没法用
qwen/qwen3-32b 总体感觉不如30b-a3b好	9K	3.62 tok/sec • 4390 tokens • 24.33s to first token • Stop reason: EOS Token Found	输出结果好
		11.23 token/s•2761 token•首个token用时 11.66 s•停止原因: 检测到 EOS token	双卡
qwen/qwen3-32b 8bit		5.04 token/s•2225 token•首个token用时 32.17 s•停止原因: 检测到 EOS token	AI MAX 395
google/gemma-3-12b	14k	2.63 tok/sec•417 tokens•304.28s to first token•Stop reason: User Stopped	错误太多
tifa-deepsex-14b-cot	7k	1.30 tok/sec •2129 tokens•96.02s to first token•Stop reason: EOS Token Found	小错误
Qwen3-8b	7K	32.84 tok/sec•2293 tokens•3.32s to first token•Stop reason: EOS Token Found	错误较多速度快呀处理token也快
MaziyarPanahi/DeepSeek-R1-0528-Qwen3-8B-GGUF/DeepSeek-R1-0528-Qwen3-8B.Q4_K_M	8K	4.38 tok/sec•3039 tokens•40.89s to first token•Stop reason: EOS Token Found	计算错误，勉强
microsoft/phi-4-reasoning-plus	16K	17.64 tok/sec•10947 tokens•5.63s to first token•Stop reason: EOS Token Found	输出结果好，速度不错
microsoft/phi-4-mini-reasoning	8K	15.83 tok/sec•3698 tokens•2.01s to first token•Stop reason: EOS Token Found	答非所问
deepseek-r1-distill-qwen-14b	6K	1.40 tok/sec•1267 tokens•72.01s to first token•Stop reason: User Stopped
bartowski/DeepSeek-R1-Distill-Qwen-14B-GGUF	6K	1.74 tok/sec•859 tokens•71.33s to first token•Stop reason: User Stopped
mistralai/mistral-7b-instruct-v0.3	8K	38.04 tok/sec •1008 tokens•4.75s to first token•Stop reason: EOS Token Found	速度快回答简单但是没有错误
Hack337/ChatGPT-5-Q8_0-GGUF/chatgpt-5-q8_0.gguf	6K	74.52 tok/sec•835 tokens•0.77s to first token•Stop reason: EOS Token Found	响应速度飞快，英语好，计算能力几乎为0
deepseek-r1-distill-llama-70b	5K	2.47 tok/sec•779 tokens•272.53s to first token•Stop reason: EOS Token Found	回答很差，半半拉拉
qwen/qwen3-30b-a3b 4bit	9K	24.18 tok/sec•4040 tokens•5.62s to first token•Stop reason: EOS Token Found	输出最好的输出速度感觉不错
		33.35 token/s•4239 token•首个token用时 48.21 s•停止原因: 检测到 EOS token	AI MAX 395 volkun
qwen/qwen3-30b-a3b 8bit		27.14 token/s•3425 token•首个token用时 50.99 s•停止原因: 检测到 EOS token	AI MAX 395 volkun
qwen3-30b-a3b@q8_0	7k	7.18 tok/sec•2370 词元•23.28 到第一个词元•停止原因: 发现EOS词元	没有错误，结果还行
gemma-3n-e4b-it-text	13K	31.42 token/s 1303 token•首个token用时 4.37 s•停止原因: 检测到 EOS token	结果不行，双卡
google/gemma-3-27b		15.32 token/s•1593 token•首个token用时 10.46 s•停止原因: 检测到 EOS token	结果还行，没有错误，双卡
llama4-dolphin-8b		39.72 token/s•154 token•首个token用时 4.60 s•停止原因: 检测到 EOS token	答非所问
goedel-prover-v2-8b@f16		8.95 token/s•2586 token•首个token用时 21.65 s•停止原因: 检测到 EOS token	答非所问
goedel-prover-v2-8b@q8_0		15.51 token/s•2715 token•首个token用时 10.00 s•停止原因: 检测到 EOS token	答非所问

养老古

不同LLM分析SDC得效果