@dh7

Hello world: a capsule that serves its own documentation. No model, no GPU, no network: a chat endpoint that greets you back, its docs for people at /human/ and for agents at /agent. Start here.

Hello world: a capsule that serves its own documentation. No model, no GPU, no network: a chat endpoint that greets you back, its docs for people at /human/ and for agents at /agent. Start here.

An image model in a capsule: Kromoss with its 8-step DMD2 LoRA on diffusers, text to 1024x1024 PNG in about 19 s on a GB10. Shows the image.generate kind, [[artifacts]] for 40 GB of weights, [params] and an extra route.

A small local model for opencode: Qwen3 4B (Q4_K_M GGUF) on llama.cpp's server, an OpenAI-compatible chat endpoint with tool calls, for opencode running on your own machine. About 6 GB of VRAM, 32k context. Needs no network.

opencode inside a capsule: the opencode coding agent in a sandbox, on DeepSeek V4.1 Flash through OpenRouter. Each session gets an empty directory where the agent can write and run code. Shows a Dockerfile, the agent.session kind and [runtime.app]. No GPU.

OpenRouter with a secret: a chat endpoint that forwards to a model on OpenRouter. Shows [network], [secrets] and [[secrets.inject]]: your OPENROUTER_API_KEY stays in m64's gate and never enters the container. No GPU.

opencode terminal on DeepSeek V4.1 Flash (OpenRouter). opencode 1.18.23's own TUI, running in tmux inside the sandbox and served to the browser by ttyd: open the app port and use opencode as in a terminal. The session survives closing the tab. Drives DeepSeek V4.1 Flash through OpenRouter. No GPU; the only egress is openrouter.ai:443, and the m64 gate adds the OpenRouter key, so the agent never holds it.

opencode agent on DeepSeek V4.1 Flash (OpenRouter). opencode 1.18.23 coding agent driving DeepSeek V4.1 Flash through OpenRouter. Each session gets an empty directory; the agent can write and run code there. No GPU; the only egress is openrouter.ai:443, and the m64 gate adds the OpenRouter key, so the agent never holds it.

opencode agent on Qwen3.8-Flash-Next (Basalt, RTX 5090). opencode 1.18.23 coding agent driving Qwen3.8-Flash-Next (IQ3_S pack, 384k context) on the Basalt engine, both in one container on one RTX 5090 with 64 GB of RAM. Each session gets an empty directory; the agent can write and run code there. opencode itself (its terminal interface, in the browser, opened on the folder home) is on a second port. Needs no network (the model runs in the same container); serve it with --allow-external-network to let the agent install packages.

Qwen3.8-Flash-Next on Basalt (RTX 5090), for opencode. A local OpenAI-compatible chat endpoint for Qwen3.8-Flash-Next (IQ3_S pack) on the Basalt engine, with a 384k context (YaRN x1.5), for opencode running on your own machine: opencode and its file access stay there, only the model calls come here. Streams, tool calls, images. About 300 tok/s on one RTX 5090 with 64 GB of RAM. Needs no network.

SANA 1.5 1.6B text-to-image on GB10 (diffusers). SANA 1.5 1.6B at 1024x1024, 20 steps, guidance 4.5, bf16. A few seconds per image on a DGX Spark (GB10).

Kromoss + DMD2 8-step LoRA text-to-image on GB10 (diffusers). lodestones/Kromoss (MOSS-Image fine-tune) with its 8-step DMD2 LoRA: 1024x1024, 8 steps, CFG 1, shift 5, bf16. About 19 s per image on a DGX Spark (GB10).

DeepSeek V4.1 Flash chat endpoint (OpenRouter). An OpenAI-compatible chat endpoint for DeepSeek V4.1 Flash on OpenRouter, for opencode (or any client) running on your own machine: opencode and its file access stay there, only the model calls come here. Streams. No GPU; the only egress is openrouter.ai:443, and the m64 gate adds the OpenRouter key, so the capsule never holds it.