Backed byY Combinator
conifer

TowardsPrivate Superintelligence

An ASCII luminance dither of a conifer, framed by a sun and moon. Hover with a pointer to ripple the characters along the row.
Backed byY Combinator
  • The Wall Street Journal.
  • The Hill
  • Liquid AI
  • NVIDIA
  • Qwen
  • Alibaba Cloud

Routing from the bottom up.

Local-first

Every query starts on your own machine and only moves up when it has to:

  • Your machine is the first stop. The weights stay loaded in unified memory, so most queries run at full speed and cost nothing.
  • A query goes to the cloud only when the local answer falls short. Cheaper cloud models are tried before the frontier ones.
  • You pay only for the queries that reach the cloud. Most never leave the machine, which cuts total token cost by more than 80%.
  • Turn on secure routing and nothing leaves the machine. The router still picks the best local model for the task.

Local inference
shouldn't feel slow.

GPU accelerated

The speed lives in the kernels. We tune the GPU code to each model and each chip:

  • Hand-written Metal and CUDA kernels on the hot paths, not a generic BLAS call.
  • Fused kernels: quantized matmul, RMSNorm, RoPE and the attention softmax run in one pass, so the weights cross the memory bus once.
  • Decode is memory-bound, so we keep weights in Q4_K and move the fewest bytes per token.
  • A flash-style attention kernel sized to each model's head dimension keeps the KV cache on-chip.
  • For mixture-of-experts models, routing is one grouped GEMM instead of a loop over experts.

The chip spends its time computing, not waiting on memory.

download · conifer

Run it on your machine.

Pick your platform. One app, one account, and the free tier is already in the box.

macOS
download
Universal · macOS 11+
Windows
download
x86-64 · Windows 10+
Linux
download
x86-64 · AppImage (.deb also available)