explainx / blog
Step-by-step: Ollama, llama.cpp, LM Studio, or vLLM → OpenAI-compatible API → OpenCode opencode.jsonc. Model picks (Qwen 3.6, GLM, Gemma), hardware tiers, multi-model routing, and troubleshooting.
Jun 29, 2026
After HN front-page hype, hands-on tests say Qwen 3.6 27B dense is the local sweet spot — better code than the 35B MoE, runnable at Q8 on 48GB RAM. Full llama.cpp + OpenCode config inside.
Jun 27, 2026
Meta Llama 4 is the most significant open-weight model release in AI history. Scout fits on a single consumer GPU with a 10 million token context window. Maverick outperforms GPT-4o on key benchmarks. And Behemoth — the 2-trillion-parameter teacher — is rewriting what we thought was possible outside a closed lab. This is your complete guide to the Llama 4 family.
Jun 19, 2026
OpenCode lets you connect any model from any provider — Claude, GPT, Gemini, GLM-5.2, local Ollama — in terminal, desktop, or IDE. With LSP auto-loading, parallel sessions, share links, and optional OpenCode Zen curated models, it is one of 2026's default open-source harnesses. How it maps to agent harness design and when to pick it over Pi or Claude Code.
July 1, 2026: With Fable 5 still offline, GLM-5.2 and Qwen 3.6 27B proving local coding is practical, the missing piece for many developers is not which model — it is how to wire inference into a harness.
OpenCode is the open-source agent loop that accepts any OpenAI-compatible API. Run weights on your CPU/GPU, expose http://127.0.0.1:…/v1, point ~/.config/opencode/opencode.jsonc at it, and you have a local coding agent — same tools, LSP, and /init AGENTS.md flow as cloud setups, without sending repo context to a third party.
This is explainx.ai's end-to-end stack guide: pick a model → start a server → configure OpenCode → tier local vs cloud.
| Step | What to do | Deep dive |
|---|---|---|
| 1. Hardware | 32GB+ for Q4 7B–27B; 48GB+ for Q8 27B | Mac vs GPU guide |
| 2. Model | Qwen 3.6 27B coding default; GLM-5.2 if you have GPU headroom | Qwen local guide · GLM Unsloth |
| 3. Runtime | llama.cpp (control) or Ollama (ease) | What is llama.cpp? · Personal AI system |
| 4. Server | OpenAI-compatible /v1 on localhost | Sections below |
| 5. OpenCode | opencode.jsonc provider + opencode in repo | OpenCode harness guide |
| 6. Verify | /init, small coding task, check tok/s | Quantization guide |
┌─────────────────────────────────────┐
│ OpenCode (agent harness) │
│ tools · LSP · AGENTS.md · sessions │
└──────────────┬──────────────────────┘
│ OpenAI-compatible HTTP
┌──────────────▼──────────────────────┐
│ Inference server │
│ llama.cpp · Ollama · LM Studio · vLLM│
└──────────────┬──────────────────────┘
│ GGUF / safetensors
┌──────────────▼──────────────────────┐
│ Open-weight model on your disk/GPU │
│ Qwen · GLM · Gemma · DeepSeek · … │
└─────────────────────────────────────┘
OpenCode never loads weights itself — it only talks HTTP. That separation is why one config file can swap Ollama today and llama.cpp tomorrow.
curl -fsSL https://opencode.ai/install | bash
# or: npm install -g opencode-ai
# or: brew install anomalyco/tap/opencode
cd your-project
opencode
First-run inside TUI:
/connect — for cloud providers (Z.AI, OpenRouter, Copilot OAuth)/init — generate AGENTS.md project memory/models — pick active model after providers existSlash reference: OpenCode commands. Harness concepts: What is an agent harness?.
| Runtime | Best for | OpenAI API default | OpenCode fit |
|---|---|---|---|
| llama.cpp | Max control, MTP, Apple Silicon | http://127.0.0.1:8080/v1 | Recommended for serious local coding |
| Ollama | One-command pull/serve | http://127.0.0.1:11434/v1 | Fastest onboarding |
| LM Studio | GUI + local server toggle | http://127.0.0.1:1234/v1 | Non-terminal users |
| vLLM | Multi-user / production GPU box | custom port /v1 | Team LAN server |
Codex users: same local servers work in Codex OSS mode — different harness, identical inference layer.
From the Qwen 3.6 local benchmark post:
llama-server -hf unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0 \
--spec-type draft-mtp -ngl 999 -fa on -c 65536 --port 8080
Smoke test: open http://127.0.0.1:8080 or curl http://127.0.0.1:8080/v1/models.
ollama pull qwen3.6:27b # verify tag on ollama.com/library
ollama serve # usually auto-starts on 11434
Other coding pulls teams use: glm-5.2, deepseek-coder-v2, gemma4:12b — check VRAM before pulling 70B-class tags.
python -m vllm.entrypoints.openai.api_server \
--model Qwen/Qwen3.6-27B-Instruct \
--port 8000
Point OpenCode at http://127.0.0.1:8000/v1 or your LAN IP for a shared team box.
opencode.jsonc)Config path: ~/.config/opencode/opencode.jsonc (user-wide) or project-local per OpenCode docs.
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"llama": {
"name": "llama.cpp (local)",
"npm": "@ai-sdk/openai-compatible",
"options": {
"baseURL": "http://127.0.0.1:8080/v1",
"apiKey": "local"
},
"models": {
"qwen3.6-27b": { "name": "Qwen3.6-27B Q8 +MTP" }
}
}
},
"model": "llama/qwen3.6-27b"
}
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"ollama": {
"name": "Ollama (local)",
"npm": "@ai-sdk/openai-compatible",
"options": {
"baseURL": "http://127.0.0.1:11434/v1",
"apiKey": "ollama"
},
"models": {
"qwen3.6-27b": { "name": "qwen3.6:27b" }
}
}
},
"model": "ollama/qwen3.6-27b"
}
"options": {
"baseURL": "http://127.0.0.1:1234/v1",
"apiKey": "lm-studio"
}
Restart OpenCode after editing config, or run opencode fresh from the project directory.
Post-Fable tiering:
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"llama": {
"name": "Local Qwen",
"npm": "@ai-sdk/openai-compatible",
"options": {
"baseURL": "http://127.0.0.1:8080/v1",
"apiKey": "local"
},
"models": {
"qwen3.6-27b": { "name": "Qwen3.6-27B local" }
}
},
"zai": {
"name": "GLM Coding Plan",
"options": {
"baseURL": "https://api.z.ai/api/coding/paas/v4"
}
}
},
"model": "llama/qwen3.6-27b"
| Workload | Switch to |
|---|---|
| Private repo, offline flight | llama/qwen3.6-27b |
| Hard refactor, long agent | /models → GLM-5.2 (harness guide) |
| Cheap volume API | Cline $9.99 open weights or OpenRouter |
Connect Z.AI via /connect inside TUI if you prefer not to store API keys in jsonc.
| Model | Local fit | OpenCode notes |
|---|---|---|
| Qwen 3.6 27B dense | Best balance 48GB Mac / 24GB+ Nvidia | Beats MoE 35B A3B on instruction following |
| Gemma 4 12B/31B | Multimodal + Apache 2.0 | Slower than Qwen for pure code |
| GLM-5.2 Unsloth | Workstation / multi-GPU | Closest to frontier local |
| DeepSeek V4 Flash (quant) | High tok/s if RAM allows | Aggressive quants trade quality |
| Kimi K2.7 | Usually API — too large for most locals | Use API via OpenCode /connect |
Alibaba cloud line: Qwen 3.7-Max for when local is not enough.
| Machine | Realistic local model |
|---|---|
| 32GB Apple Silicon | Q4 27B or Q8 7B–14B |
| 48–64GB Apple Silicon | Q8 Qwen 3.6 27B (~32 tok/s MTP) |
| 24GB Nvidia (4090/5090) | Q6 27B; Q4 70B tight |
| 64GB+ VRAM / dual GPU | GLM-5.2 class, vLLM team server |
Full economics: Mac vs dedicated GPU · closed vs open cost table.
curl -s http://127.0.0.1:8080/v1/models (or Ollama 11434)~/.config/opencode/opencode.jsoncopencode in git repo/init — creates AGENTS.mdAdd verification loops from explainx.ai loops — e.g. ci-until-green.
| Symptom | Fix |
|---|---|
| Connection refused | Inference server not running or wrong port in baseURL |
| Empty / garbage output | Quant too low — bump Q4→Q6/Q8 (quant guide) |
| Slow first token | Model loading; MTP helps decode — see llama.cpp --spec-type draft-mtp |
| Ignores package.json / structure | Try dense over MoE; shorten context if RAM swapping |
| Tool calls fail | Model may lack reliable function-calling — switch to GLM API or GPT-class for agent-heavy runs |
| Wrong provider | echo config path; /models list; check model string matches provider/id |
| Harness | Local OSS models | Notes |
|---|---|---|
| OpenCode | Yes — 75+ providers + custom local | Default open multi-surface |
| Pi | Yes — provider plugins | Minimal, ownable |
| Codex OSS | Yes — --oss + Ollama | OpenAI-native tool schema |
| Claude Code | Indirect — /config remote to desktop | Not weight-local |
| Kilo Code | Yes — BYOK + Ollama | VS Code extension |
OpenCode wins when you want one harness, terminal + desktop, and swap local/cloud without reinstalling.
Official: OpenCode · OpenCode config schema · llama.cpp · Ollama
Runtime ports, model tags, and OpenCode provider schema reflect July 1, 2026 docs — verify before production. Last updated: July 1, 2026.