I build and test local AI.
Articles about AI software and language models
I'm Django de Vreng. I build local AI systems, run open models on the DGX Spark and write about my work here. I share the test settings with my measurements and explain what the results mean.
- 38
- Model profiles
- 10
- Different tests
- 380
- Measurements
- 128 GB
- Memory in my DGX Spark
Gemma-4 on the DGX Spark: NVFP4 compared with BF16
These five tests compare decode speed per user with ten concurrent requests. Bars show the percentage difference of NVFP4 relative to BF16, calculated from the means in the Arena.
This compares two configurations with different model and kernel settings. These measurements do not assess answer quality.
Chat: 1024 input and 1024 output tokens, c=10. BF16: vLLM v0.26.0. NVFP4: vLLM v0.26.0.
View both profiles →Difference in decode speed
All bars: 0–125% difference. Length shows the size of the difference; + means faster, − means slower. A dash means a comparable measurement is missing.
What can you run locally, and what do you need?
Compare language models
See model speed and memory use on my DGX Spark. The test settings are included with the results.
Compare models DGX SparkRun models yourself
Read how I configure vLLM, fit models into memory and test multiple concurrent requests.
Open the DGX Spark guide Local AIWhen should you choose local?
What does running locally mean for costs, data and maintenance? I compare it with models offered through a cloud service.
Open the decision guide-
18-08-26·Build logs001Build logs
Making my blog readable for AI agents
I added Markdown versions, an agent index, and build checks to every post. Not as a GEO trick, but to make the underlying source less ambiguous.
-
23-06-26·On-prem AI002On-prem AI
Gemma-4 v23 on the DGX Spark
New vLLM v0.23.0 runs for Gemma-4 on the DGX Spark: BF16, NVFP4 and MTP compared across decode, TTFT, tails and practical local-agent limits.
-
22-05-26·On-prem AI003On-prem AI
The three numbers behind a fast DGX Spark
Decode, prefill and queueing: three numbers decide whether a DGX Spark feels fast under a real workload, and those three are exactly what most reviews skip.
-
05-05-26·Reflections004Reflections
Why this blog and arena exist
I looked for concrete numbers on local AI on the DGX Spark and never found them. So I measure them myself, building the blog and arena as an open workbench.
-
03-05-26·On-prem AI005On-prem AI
Gemma-4 on the DGX Spark: NVFP4 vs BF16
Nine identical benchmarks, two precisions. NVFP4 runs 22 to 92 percent faster per token, and peak-hour capacity grows 69 percent on the Spark.
I'm Django.
I build the layer between language models and real work: agents, MCP servers, local models and on-prem AI. They let language models retrieve documents and use other software. I also work on error handling: what happens if a tool does not respond, for instance?
Here I write about those projects and my tests on the DGX Spark. I explain my choices and the problems I encounter. Many articles come from my work at Kamoo. This is my personal blog.
More about me →