puding 🥣 — model results

LLM benchmark on a GB10 (Nvidia Spark GX10, 128GB unified memory). Each model gets the same coding prompts via an OpenAI-compatible API; I record speed and token usage, and keep the generated artifact (an SVG or a single-file app) so you can open it below. Github: https://github.com/polaroi8d/puding

Ornith-1.0-35B

Prompt	TTFT	Total Time	Prompt Tok	Completion Tok	Total Tok	Speed
SVG Pelican on a Bicycle	0.381s	106.251s	65	3168	3233	29.8 tok/s
Todo App	0.702s	77.964s	1122	2305	3427	29.6 tok/s

Qwen/Qwen-AgentWorld-35B-A3B

Prompt	TTFT	Total Time	Prompt Tok	Completion Tok	Total Tok	Speed
SVG Pelican on a Bicycle	1.442s	528.341s	65	15552	15617	29.4 tok/s
Todo App	1.923s	336.832s	1122	9909	11031	29.4 tok/s

Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8

Prompt	TTFT	Total Time	Prompt Tok	Completion Tok	Total Tok	Speed
SVG Pelican on a Bicycle	0.241s	16.496s	63	900	963	54.6 tok/s
Todo App	0.645s	29.208s	1082	1525	2607	52.2 tok/s

Qwen/Qwen3.6-35B-A3B

Prompt	TTFT	Total Time	Prompt Tok	Completion Tok	Total Tok	Speed
SVG Pelican on a Bicycle	1.501s	264.784s	65	7720	7785	29.2 tok/s
Todo App	2.129s	87.112s	1122	2560	3682	29.4 tok/s

Qwen3-Coder-Next-FP8

Prompt	TTFT	Total Time	Prompt Tok	Completion Tok	Total Tok	Speed
SVG Pelican on a Bicycle	0.386s	38.387s	63	1822	1885	47.5 tok/s
Todo App	0.844s	50.049s	1082	2346	3428	46.9 tok/s

mistralai/Codestral-22B-v0.1

Prompt	TTFT	Total Time	Prompt Tok	Completion Tok	Total Tok	Speed
SVG Pelican on a Bicycle	0.835s	80.331s	65	431	496	5.4 tok/s
Todo App	1.117s	61.132s	1254	324	1578	5.3 tok/s

openai/gpt-oss-20b

Prompt	TTFT	Total Time	Prompt Tok	Completion Tok	Total Tok	Speed
SVG Pelican on a Bicycle	2.99s	12.262s	124	511	635	41.7 tok/s
Todo App	9.476s	36.043s	1146	1613	2759	44.8 tok/s

qwen38-27b-unsloth-nvfp4

Prompt	TTFT	Total Time	Prompt Tok	Completion Tok	Total Tok	Speed
SVG Pelican on a Bicycle	1534.338s	1735.15s	103	31595	31698	18.2 tok/s
Todo App	773.128s	1007.585s	1160	19757	20917	19.6 tok/s