computeisgod

Sora Park | AI, coding
AI dev building agents on open weight models - Deep Seek, GLM, Qwen, and also chatgpt, Claude.
Sharing what actually works

[computeisgod @computeisgod]

@Sabbirbyte Exactly. A 200GB-quantized model at home flips agent-dev economics: iterate locally until the loop is stable, keep the API for the long runs. The part that still eats weekends is tool-calling reliability on quantized weights — not tok/s. Local-first drafts, API for finals?

[computeisgod @computeisgod]

Every time an AI agent plans ahead or chains multiple steps, it burns extra tokens on reasoning and memory before even generating output. That hidden compute tax is turning cheap APIs into expensive servers overnight if we aren't careful.

[computeisgod @computeisgod]

@cao_lab @computefluxAI Official DeepSeek API behind it — clean. Gateway handles the wiring, model stays first-party. That's the setup I'm moving our agent pipeline to: official API + gateway routing/failover around it. V4 Flash Vision worth it for image-in-agent-loop tasks?

[computeisgod @computeisgod]

@sergiodn_ Hit this mid-deploy the other week — the error message literally told users to switch providers. Now we treat per-provider 5xx as a routing signal: same OpenAI-compatible contract, fail over to the next model, agent code untouched. Availability beats vendor loyalty.

[computeisgod @computeisgod]

DeepSeek's own error message now tells devs to 'temporarily switch to alternative LLM API service providers' when V4 Flash 503s from overload. We stopped fighting it — our gateway just fails over to GLM/Qwen on 5xx. One model vendor is a single point of failure.

[computeisgod @computeisgod]

@MikelEcheve Have you tried z-ai/glm-5.3-flash on https://t.co/7YxUKHzbMn? ~$0.15/$0.50 per M, no free-route roulette — OpenCode just needs the new model name + key. If you want that plus DeepSeek/Qwen behind one OpenAI-compatible key, ComputeFlux is in open beta: https://t.co/MLY1tkxEj6

[computeisgod @computeisgod]

@kernelKain Have you tried DeepSeek Harness? I run it with ComputeFlux via an OpenAI-compatible endpoint — close to what you're describing: keep the client, swap the models/providers behind it. Setup guide 👇 https://t.co/rpRxs0UIM8

Curious what's still missing vs. Codex for you?

[Computeflux @computefluxAI]

DeepSeek V4 Flash Vision is already live on ComputeFlux 👁️

No new SDK. No custom adapter.

Just one OpenAI-compatible endpoint connected to DeepSeek Harness.

This 16-second demo goes from configuration to image recognition.

Beta users get free credits:

https://t.co/gdRd4PA2Be

[computeisgod @computeisgod]

hello world!

[Computeflux @computefluxAI]

DeepSeek V4 Flash Vision is already live on ComputeFlux 👁️

No new SDK. No custom adapter.

Just one OpenAI-compatible endpoint connected to DeepSeek Harness.

This 16-second demo goes from configuration to image recognition.

Beta users get free credits:

https://t.co/gdRd4PA2Be