Announcement
Conifer: one endpoint, routing decides.
You call https://api.conifer.build. Name a model and that model is served. Use a route or reach preset and everyday work can stay on your machine. Mint a key in the console.
Models, providers, and tools arrive separately. Conifer is the layer that holds them together: one OpenAI-compatible and Anthropic-compatible endpoint, a router that picks the venue, and a console for the key and the bill.
Local weights stay on your hardware. Cloud models sit behind the same interface, through a Conifer key or a key you already have. A named model is that model. A route decides the rest.
Why build this
Open models already cover most everyday work. The leftover is setup, and the hard fifth of requests that still need a frontier model. Conifer’s job is to make both cases one call.
What’s included
One system, not a kit to assemble:
- The gateway at
https://api.conifer.build, on the OpenAI and Anthropic wires. - The console, where you mint a key, add your own provider keys, and read spend.
- Cloud routing, on a saved route or reach preset: easy work stays local, a frontier model is paid for only when the request needs one.
- The CLI: full engine on Apple Silicon and Windows. Linux stays a thin client.
- A signed desktop app on macOS. Windows and Linux desktop pages stayed coming soon. (The download page was retired 2026-09-01.)
Benchmarks
The engine is benchmarked head-to-head against the field on the same hardware, with the same models and workloads, and nothing leaving the device.
On an Apple M3 Max with the Metal backend, decode runs close to the memory-bandwidth wall, the limit that ultimately caps local token generation, reaching up to about 89% of the chip’s theoretical bandwidth on the 7–8B models. Against llama.cpp on identical Q4_K_M weights, Conifer leads on decode across every model tested and sits at parity on prefill. MLX, Apple’s own framework, still leads on decode for several models; Conifer is ahead on prefill for many.
| Model | Conifer | llama.cpp |
|---|---|---|
| LFM2-350M | 764 | 506 |
| Llama-3.2-1B | 270 | 206 |
| Qwen2.5-7B | 57 | 53 |
| Llama-3.1-8B | 55 | 51 |
| Gemma-3-12B | 26 | 24 |
Apple M3 Max (36 GB), Metal backend, Q4_K_M weights, 512-token prompt and 128-token decode, best of three runs. Per-model figures live in the model ledger.
A few highlights from the same run:
- Fastest decode: 764 tok/s on LFM2-350M.
- Lowest energy: 0.06 joules per token on LFM2-350M.
- Best efficiency: about 16 tokens per second per watt on LFM2-350M.
- Local marginal cost: from $0.0031 per million tokens: electricity, not an API bill, with your data staying on-device.
On Windows with NVIDIA, the CUDA backend reaches parity with llama.cpp on decode; CUDA prefill and the Vulkan backend are still being optimized.
Why local
Cloud AI is increasingly priced and scaled for the frontier, which is more than most everyday work needs. Open-weight models now cover that work. You should not have to weigh cloud against local for every task. Conifer defaults to on-device execution, fast, private, and predictable, and in privacy mode your prompts, documents, and outputs never leave the machine in front of you.
What ships today
The gateway, the key, the console, and cloud routing are live. They are not a later phase.
Quick start
- Mint a key at the console.
- Point a client at
https://api.conifer.build. - Or install the CLI:
curl -fsSL https://conifer.build/setup | bash(Windows: the PowerShell line on Platforms). - Optional: the signed macOS app (its download page was retired 2026-09-01).