Post History

Current version by Nick Antonaccio

Current VersionJun 25, 2026 at 11:00

It should be noted that Deepseek V4 Flash doesn't currently run in LM Studio, Llama CPP, vLLM, or other common inference software, because it's architecture is different than other models, but everything you need to run it locally is here:

https://github.com/antirez/ds4

I can confirm that the q2-imatrix version of Deepseek V4 Flash runs within the ds4 software from that repository, on a single DGX Spark (Asus GX10) at about 10 tokens per second, with 300k context size.

I told Pi to install and run it (and the entire installation task used V4 Flash as the provider model on Openrouter, haha :) ), and it did everything without a hitch, including evaluating which quant to use, downloading the model, compiling and running the ds4 software, wiring up the new server as an available provider model in Pi, and documenting everything needed to install the whole thing on other machines.

Tomorrow I'll try the q4-imatrix quant of V4 Flash on 2 clustered DGX Spark machines. It should be a perfect fit, with enough VRAM left over for 1 million token context length, and excellent inference quality (the ds4 quants are quite different than other typical quantizations, because some important layers are compressed less than others, providing better accuracy where needed).

V4 Flash in ds4 certainly isn't the fastest model in my local stable, but I don't need to run it locally for most work. I love that I can use it for what feels like almost free, over the API from Deepseek directly, from Openrouter, Ollama, and plenty of other providers. And for clients who need HIPAA compliant in-house inference, it feels very close to frontier quality, on relatively inexpensive self-hosted hardware which doesn't require a new electrical panel installation.

Previous Versions
Version 5Jun 25, 2026 at 11:00

It should be noted that Deepseek V4 Flash doesn't currently run in LM Studio, Llama CPP, vLLM, or other common inference software, because it's architecture is different than other models, but everything you need to run it locally is here:

https://github.com/antirez/ds4

I can confirm that the q2-imatrix version of Deepseek V4 Flash runs within the ds4 software from that repository, on a single DGX Spark (Asus GX10) at about 10 tokens per second, with 300k context size.

I told Pi to install and run it (using V4 Flash is the provider model on Openrouter, haha :) ), and it did everything without a hitch, including evaluating which quant to use, downloading the model, compiling and running the ds4 software, wiring up the new server as an available provider model in Pi, and documenting everything needed to install the whole thing on other machines.

Tomorrow I'll try the q4-imatrix quant of V4 Flash on 2 clustered DGX Spark machines. It should be a perfect fit, with enough VRAM left over for 1 million token context length, and excellent inference quality. It certainly isn't the fastest model in my local stable, but I love that I can use it for what feels like almost free, over the API from Deepseek directly, from Openrouter, Ollama, and plenty of other providers. And for clients who need HIPAA compliant in-house inference, it feels very close to frontier quality, on relatively inexpensive self-hosted hardware which doesn't require a new electrical panel installation.

Version 4Jun 25, 2026 at 10:55

Deepseek V4 Flash doesn't currently run in LM Studio, Llama CPP, vLLM, or other common inference software, because it's architecture is different than other models, but everything you need to run it locally is here:

https://github.com/antirez/ds4

I can confirm that the q2-imatrix version of Deepseek V4 Flash runs within the ds4 software from that repository, on a single DGX Spark (Asus GX10) at about 10 tokens per second, with 300k context size.

I told Pi to install and run it (using V4 Flash is the provider model on Openrouter, haha :) ), and it did everything without a hitch, including evaluating which quant to use, downloading the model, compiling and running the ds4 software, wiring up the new server as an available provider model in Pi, and documenting everything needed to install the whole thing on other machines.

Tomorrow I'll try the q4-imatrix quant of V4 Flash on 2 clustered DGX Spark machines. It should be a perfect fit, with enough VRAM left over for 1 million token context length, and excellent inference quality. It certainly isn't the fastest model in my local stable, but I love that I can use it for what feels like almost free, over the API from Deepseek directly, from Openrouter, Ollama, and plenty of other providers. And for clients who need HIPAA compliant in-house inference, it feels very close to frontier quality, on relatively inexpensive self-hosted hardware which doesn't require a new electrical panel installation.

Version 3Jun 22, 2026 at 02:48

Deepseek V4 Flash doesn't currently run in LM Studio, Llama CPP, vLLM, or other common inference software, because it's architecture is different than other models, but everything you need to run it locally is here:

https://github.com/antirez/ds4

I can confirm that the q2-imatrix version of Deepseek V4 Flash runs within the ds4 software from that repository, on a single DGX Spark (Asus GX10) at about 10 tokens per second, with 300k context size.

I told Pi to install and run it (using V4 Flash is the provider model on Openrouter, haha :) ), and it did everything without a hitch, including evaluating which quant to use, downloading the model, compiling and running the ds4 software, wiring up the new server as an available provider model in Pi, and documenting everything needed to install the whole thing on other machines.

Tomorrow I'll try the q4-imatrix quant of V4 Flash on 2 clustered DGX Spark machines. It should be a perfect fit, with enough VRAM left over for 1 million token context length, and excellent inference quality. It certainly isn't the fastest model in my local stable, but I love that I can use it for what feels like almost free, over the API from Deepseek directly, from Openrouter, Ollama, and plenty of other providers. And for clients who need HIPAA compliant in-house inference, it feels very close to frontier quality.

Version 2Jun 22, 2026 at 02:46

You can't currently run Deepseek V4 Flash in LM Studio, Llama CPP, vLLM, or other common inference software, but everything you need to run it locally is here:

https://github.com/antirez/ds4

I can confirm that the q2-imatrix version runs in the ds4 software from that repository, on a single DGX Spark (Asus GX10) at about 10 tokens per second, with 300k context size.

I just told Pi to install and run it, and it did everything without a hitch, including evaluating which quant to use, downloading the model, compiling and running the ds4 software, wiring up the server as an available provider model in Pi, and documenting everything needed to install on other machines.

Tomorrow I'll try the q4-imatrix quant on 2 clustered DGX Spark machines. It should be a perfect fit with enough VRAM left over for 1 million token context length, and excellent inference quality :)

Version 1Jun 22, 2026 at 01:56

You can't run Deepseek V4 Flash in LM Studio, Llama CPP, vLLM, or other common inference software, but everything you need to run it locally is here:

https://github.com/antirez/ds4

I can confirm that the q2-imatrix version runs in the ds4 software from that repository, on a single DGX Spark (Asus GX10) at about 10 tokens per second, with 300k context size.

I just told Pi to install and run it, and it did everything without a hitch, including evaluating which quant to use, downloading the model, compiling and running the ds4 software, wiring up the server as an available provider model in Pi, and documenting everything needed to install on other machines.

Tomorrow I'll try the q4-imatrix quant on 2 clustered DGX Spark machines. It should be a perfect fit with enough VRAM left over for 1 million token context length, and excellent inference quality :)