Submit a recipe
Every recipe enters the repo as a pull request and goes live after validation and a manual review.
Ready?
Fork the repo and open a PR. Not comfortable with JSON? Open an issue with the source link instead.
Steps
- Fork the repo and create a branch
- Append one entry to the array in
data/recipes.json(JSON; see the template below andschema/) - Run
node scripts/validate.mjsandnode scripts/check_dupes.mjslocally - Open a PR with the original post / log link; for your own runs attach a screenshot or the full command output
- A maintainer checks the source level and merges; READMEs are generated by scripts, don't edit them by hand
JSON template
{
"id": "text-qwen38-27b-m4max-omlx",
"category": "text",
"model": "Qwen3.8-27B",
"model_family": "Qwen",
"quant": "UD-Q4_K_XL",
"engine": "oMLX",
"engine_note": "oMLX 0.7",
"hardware": "Mac Studio M4 Max 128GB",
"hardware_family": "Mac",
"hardware_cost_usd": 3499,
"hardware_cost_text": "$3,499 (Apple Store, 2026-10)",
"metric": "62 tok/s decode; 850 tok/s prefill (8K prompt)",
"tok_s": 62,
"concurrency": "1",
"context": "128K",
"launch_command": "omlx serve --model ... --ctx 131072",
"source_url": "https://x.com/...",
"source_note": "@author on X",
"source_level": "实测",
"date": "2026-10-13",
"tags": ["dense", "4bit"]
}Use null for unknown fields; never estimate. hardware_family: Mac | NVIDIA 消费级 | DGX Spark | AMD | 二手数据中心卡 | 其他. source_level: 官方 (official) | 实测 (measured) | 宣称 (claimed). Keep these values in Chinese in the data. tags / model_family are normalized by scripts/tag_recipes.py.
Scope
This list is for setups that one person, a home or office desktop user, or a small studio can build. Datacenter setups are out of scope. Used datacenter cards in a desktop PC are in scope. A setup is in scope only when it meets all of these conditions:
- The hardware costs $30,000 or less.
- It uses at most 4 machines. Budget is the main limit: a multi-machine setup, or a setup with more than 4 GPUs in total, is in scope when the hardware cost stays at $30,000 or less. Example: 2 machines with 4× RTX 3060 each (8 GPUs in total) are in scope when they stay under budget.
- It runs on ordinary home or office power: about 2 kW in total, from up to two ordinary outlets, with no special wiring. If you use two outlets, put them on separate circuits. One US 120 V / 15 A circuit gives about 1.44 kW continuous. One Japanese 100 V / 15 A outlet is rated at about 1.5 kW.
- It uses no rack, no SXM or HGX server board, no custom liquid-cooling loop and no server liquid cooling. A DIY SXM2-to-PCIe adapter build that fits in a desktop is in scope. A consumer GPU with a factory all-in-one (AIO) liquid cooler is in scope.
- It uses at most one desktop-class network switch.
- It uses no modded GPU. A modded GPU has changed hardware, for example changed memory. Examples are an RTX 4090D with 48GB memory and a CMP 170HX with changed memory. Recipes for modded GPUs that are already in the list stay until a later review. A stock GPU with a firmware or driver unlock is in scope, for example the gpu-unlock kernel module that removes the FP16 limit on Ampere cards. These recipes get the
gpu-unlocktag.
A setup that fails one condition is out of scope, even when it costs less than $30,000. We record the price on the date that we add the recipe. When the price goes up later, we update only the price note (hardware_cost_text). We do not remove the recipe. python3 scripts/scope_rules.py lists the recipes that fail the cost or hardware-type check. Reviewers check the machine count by reading the source.
Rules
- No estimates, nothing made up; unknown fields are
nulland show as “no public data” - Store links and summaries only; don't copy full X posts
- Source levels: official = project / vendor; measured = the poster ran it (an engine or recipe author must also publish raw benchmark data and reproduction steps); relayed = second-hand quote (not accepted for new recipes: link the original source); claimed = prediction or no reproduction details
- Inclusion focus: add new models and new hardware in the first few weeks after launch. Add non-mainstream combinations too (multi-GPU, multi-machine, AMD and Apple mixes, budget builds). Merge a new result into an existing entry only when the tester, machine, model and engine are all the same: the newest reliable value becomes the main value, and older values and sources go in the notes. A routine repeat by a different tester (a model × engine × hardware combination that is already in the list) stays a separate entry.