The right architecture, co‑designed on silicon
We design model architectures with the target hardware in the loop, scored on task quality and on-device latency and memory.
How it works
We start from the target device and runtime, and the memory and latency budget they allow.
Our Automated Foundation Model Design (AFMD) process evolves candidate architectures and scores each one on downstream task quality, plus latency and peak memory measured on the device.
We train the selected architecture for the target tasks. Data never leaves the customer’s environment.
We support OEM and on-prem deployment, with dedicated support and SLAs.
We designed LFM3 by searching across device and backend combinations
Our own model families were designed with the same hardware-in-the-loop search.
Hardware in the LFM3 search
14 device and backend combinations, up from the CPUs in one phone and one laptop for LFM2.
Samsung Galaxy S26 Ultra
OnePlus 15
Raspberry Pi 5
Apple M4 Max
AMD Ryzen AI
Arm and x86 cloud CPUs
GPUs
Qualcomm Hexagon NPU
What the search chose
LFM3 was redesigned for agentic workloads, which need more capability per parameter and fast inference over long contexts. In its 24-layer backbone, sliding-window attention gave the best quality–speed tradeoff.
In the LFM2 search, under identical on-device budgets, adding linear-attention, state-space, or additional convolution operators did not improve aggregate quality and typically worsened device metrics.
Up to 2× faster
35% fewer parameters
About 1 GB
1.6 GB KV cache
On-device performance for
Explore more quants, runtimes, modalities, and metrics on Pipette.