Lamb Labs MPU
Model Processing Unit: a custom chip whose architecture and weights are hard-coded for one model, targeting 20,000+ tokens per second for a 27B-parameter model; not yet a chip on sale, with a one-bit model (Larry) to chat with live and an FPGA prototype as proof stages.
Find alternatives to Lamb Labs MPU
What it does
- Accelerator silicon
- Model inference
Features 4
- Model in siliconThe whole model, weights included, is hard-coded into the chip.
- Speed targetAims at 20,000+ tokens per second by removing the memory-bandwidth bottleneck.
- LarryA one-bit version of Qwen3.8 27B that can be chatted with live.
- Little LambA tiny model generating text entirely inside FPGA fabric.
Sources
- Pricing
-
No pricing is published on the home page; the page lists a contact email and no way to buy the chip.
- Description
-
Official documentation · 19 Oct 2026
Home page: "Custom silicon for each model, its architecture hard-coded on-chip."
- What it does
-
Official documentation · 19 Oct 2026
Home page: "Custom silicon for each model, its architecture hard-coded on-chip."
- Sold within
-
The home page offers no way to buy a chip; it gives a contact email.
- Name
-
Official documentation · 19 Oct 2026
Name as published on the product page.
- URL
-
Official documentation · 19 Oct 2026
Product page, read 2026-10-19.
- Deployment
-
Official documentation · 19 Oct 2026
Home page: the chip runs the model on-chip with no DRAM in the token loop; the FPGA prototype runs on an off-the-shelf Kria KV260 board.
Open Lamb Labs in the directory
Something wrong here? Send a correction — quote this product id: lamb-labs-mpu.