Post History

Current version by Nick Antonaccio

Current VersionJul 04, 2026 at 01:54

UPDATE:

I've been using Deepseek V4 Flash as my daily driver ever since the first post in this topic, and haven't looked back. It's absolutely fantastic. It hasn't failed me yet in a massive pile of agentic tasks and lots of software development work.

I'm now using the cline-pass API for even less expensive inference cost, and extremely fast API performance. I keep the OpenRouter and https://api.deepseek.com APIs set up as backups.

I absolutely recommend trying cline-pass for all your open source API model inference. It's an amazing buy, especially with Deepseek V4 Flash. See https://aibynick.com/thread/51 .

For self-hosting, I've been very happy with the Antirez DS4 q2 quant of V4 Flash, which runs on a single DGX Spark machine. That quant is not a normal 2 bit compression - it leaves some select layers in the model all the way at 8 bit precision, and on a DGX Spark, it loads fully and leaves lots of room for KV Cache (I've been running with 300k cache). If I need better quality, the DS4 q4 quant runs on 2 clustered DGX Sparks, with even more headroom. It's awesome to have a locally hosted version of exactly the same model I'm using via the APIs for all my daily work.

Previous Versions
Version 3Jul 04, 2026 at 01:54

UPDATE:

I've been using Deepseek V4 Flash as my daily driver ever since the first post in this topic, and haven't looked back. It's absolutely fantastic. It hasn't failed me yet in a massive pile of agentic tasks and lots of software development work.

I'm now using the cline-pass API for even less expensive inference cost, and extremely fast API performance. I keep the OpenRouter and https://api.deepseek.com APIs set up as backups.

I absolutely recommend trying cline-pass for all your open source API model inference. It's an amazing buy, especially with Deepseek V4 Flash.

For self-hosting, I've been very happy with the Antirez DS4 q2 quant of V4 Flash, which runs on a single DGX Spark machine. That quant is not a normal 2 bit compression - it leaves some select layers in the model all the way at 8 bit precision, and on a DGX Spark, it loads fully and leaves lots of room for KV Cache (I've been running with 300k cache). If I need better quality, the DS4 q4 quant runs on 2 clustered DGX Sparks, with even more headroom. It's awesome to have a locally hosted version of exactly the same model I'm using via the APIs for all my daily work.

Version 2Jul 04, 2026 at 01:52

UPDATE:

I've been using Deepseek V4 Flash as my daily driver ever since the first post in this topic, and haven't looked back. It's absolutely fantastic. It hasn't failed me yet in a massive pile of agentic tasks and lots of software development work.

I'm now using the cline-pass API for even less expensive inference cost, and extremely fast API performance. I keep the OpenRouter and https://api.deepseek.com APIs set up as backups.

For self-hosting, I've been very happy with the Antirez DS4 q2 quant, which runs on a single DGX Spark machine, with lots of room for KV Cache. If I need better quality, the DS4 q4 quant runs on 2 clustered DGX Sparks, with even more headroom. It's awesome to have a locally hosted version of exactly the same model I'm using via the APIs for all my daily work.

Version 1Jul 04, 2026 at 00:55

UPDATE:

I've been using Deepseek V4 Flash as my daily driver ever since the first post in this topic, and haven't looked back. It's absolutely fantastic. It hasn't failed me yet in a massive pile of agentic tasks and lots of software development work.

I'm now using the cline-pass API for even less expensive inference cost, and extremely fast API performance. I keep the OpenRouter and https://api.deepseek.com APIs set up as backups.

For self-hosting, I've been very happy with the Antirez DS4 q2 quant, which runs on a single DGX Spark machine, with lots of room for KV Cache. If I need better quality, the DS4 q4 quant runs on 2 clustered DGX Sparks, with even more headroom. It's awesome to have a locally hosted version of exactly the same model I'm using via the APIs for all my daily work.