Has anyone used DeepSeek-Reasonix with a Strata Model? (DeepSeek-v4.1-Flash-IST-DAS-2.00bpw in my case)

[darrenaddy]

I have a Windows machine with 128GB RAM (i9-14900k) and a single RTX 3090Ti GPU.
Have Strata installed and running in chat with a 32K context (so far). Would DeepSeek-Reasonix be a help if I wanted to try agentic coding?

[vectornguyen76]

I haven't run this exact pairing, but it should work. Strata's DeepSeek v4.1 server exposes /v1/chat/completions and /v1/messages, and it converts the model's DSML tool calls into OpenAI tool_calls / Anthropic tool_use (Strata DEEPSEEK41.md, "Through the API"). Reasonix takes any OpenAI-compatible endpoint as a custom provider (GUIDE.md#L656-L735).

Setup in the Reasonix desktop app: Settings → Model → Access → Add model service → Custom provider

• API address: http://127.0.0.1:8080/v1. That's Strata's default address (TERMINAL_LAUNCH.md). Reasonix appends /chat/completions itself. Use your own port if you launched with --port.
• API key: if your local launch has no key, leave the key and api_key_env empty. Otherwise paste the key Strata generated.
• Model: discovery tries /models and /v1/models. If nothing shows up, type the model name the server reports.
• Context window: set it to what Strata was started with (32768 for you). Reasonix uses this number to decide when to compact. 0 turns compaction off, and on a 32K window you don't want that (GUIDE.md#L753).

Two practical caveats:

• 32K is tight for agentic coding. The system prompt, the tool schemas and a few file reads fill it quickly, so expect frequent compaction. If VRAM allows, relaunch with strata launch ... --context 64k. Larger contexts take VRAM away from the expert cache, and Strata's docs say so.
• Speed. A 3090 Ti isn't in Strata's measured table. For comparison, the RTX 3060 12 GB row shows roughly 9–11 tok/s output and ~155 tok/s prompt reading. An agent sends long prompts every step, so even with Strata's prompt caching, expect each step to take noticeably longer than a chat reply.

Reasonix isn't in strata connect's list of apps. Still, strata connect opencode --check sends one prompt and one tool call to the same OpenAI-compatible endpoint, so it's a quick way to confirm tool calling works before you point Reasonix at it.