# model-for-opencode: a small local model on an OpenAI-compatible endpoint,
# for an opencode (or any client) running outside the capsule.
#
# What this example adds to hello-world:
#   - a GPU: [hardware] says what it needs, [runtime] gpu = true;
#   - [[artifacts]]: the weights, downloaded and hashed by m64, mounted read-only.
#     Weights never go in the folder (5 MB per file, 10 MB in total);
#   - a ready-made engine image (llama.cpp's server), pinned by digest.
# It needs no network at run time and no secret: everything is local.
schema = "m64/capsule@1"
description = "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."
kinds = ["chat"]
authors = ["motif42"]
license = "mit"

[hardware]
platform = "linux-x86_64"
# Matched as a substring of the GPU's name, so any NVIDIA card passes; a
# mismatch only warns. The image below also has an arm64 build: on a GB10,
# set platform = "linux-aarch64" and gpu = "NVIDIA GB10".
gpu = "NVIDIA"
gpu_count = 1
# 2.5 GB of weights + 2.4 GB of 8-bit KV cache at 32k tokens + CUDA's overhead.
vram_gb = 6
# The image is CUDA 12.8.
driver = ">=570"
disk_gb = 6

[runtime]
sandbox = "container"
# ghcr.io/ggml-org/llama.cpp:server-cuda, build b12030 (llama.cpp commit
# a91eab67d8389d979d29fdf6633b2edd0718eb4b), multi-arch index digest. Tags move;
# a digest is exactly one image.
image = "ghcr.io/ggml-org/llama.cpp@sha256:75037a842f2ec1d1b5317ce64078bbb4a2a134238e91a755c10ce31d8f90550a"
gpu = true
internal_port = 9000
# llama-server answers 503 while it loads the model, 200 once it is ready.
health = "/health"
expected_load_seconds = 60
# The front rewrites every request's "model" to this; start.sh gives
# llama-server the same alias. It is also the name to use in opencode.
model_name = "qwen3-4b"
user = "65534:65534"

# The model: Qwen's own GGUF build, at a fixed commit. m64 downloads only the
# files in `include`, checks them, and mounts them read-only at `mount`. Pickle
# formats (.bin .pt .pth .ckpt .pkl) are refused: GGUF and safetensors only.
[[artifacts]]
type = "hf"
repo = "Qwen/Qwen3-4B-GGUF"
revision = "bc640142c66e1fdd12af0bd68f40445458f3869b"
include = ["LICENSE", "Qwen3-4B-Q4_K_M.gguf"]
mount = "/weights/qwen3-4b"