Skip to content

Give every devicean LLM of its own.

Fanshi casts large language models directly into compute-in-memory silicon: three silicon routes, one family of inference cores, one line from model to tape-out. No network, no data leaving the device, near-zero marginal cost per inference.

Watt-class handhelds
to 35B-model nodes.

The same inference cores and three silicon routes can cast any model into a chip. Pick one that fits, or have one built around your model.

Fanshi first-generation edge inference chips Forge, Argus and Trident side by side on a light grey surface
Fig. 1 · First-generation ASICs, left to right: Forge, Argus, Trident

Cast the whole model
into a single chip.

GPUs and NPUs spend most of their time and power shuttling weights between memory and compute. Fanshi turns that around: the weights live in the memory array, and the math happens where they are.

Inside the silicon / ARCHITECTURE STUDY01 / 03

Weights on the move, energy on the move.

3D preview

External weight traffic per token

800MB

Traffic comes from

Off-chip reads · cross-chip transfer

0.8B · INT8 estimate, not system power
Fig. 2 · Where the weights live decides where the power goesSchematic · not an actual layout
  1. Separate memory

    Every token moves every weight

    In a conventional NPU or GPU, memory and compute are separate pieces of silicon. For every token generated, all of the weights cross the bus: bandwidth sets the ceiling, and most of the energy is spent in transit.

  2. Compute-in-memory

    Put the weights in the memory array

    Fanshi builds the multiply-accumulate circuits into the memory array itself. Weights stay in the array, activations flow in, results flow out. No weights on the bus.

  3. MaskROM

    Cast at tape-out, never moved at runtime

    Mask ROM compute-in-memory writes the weights physically into the via array during tape-out. The chip leaves the fab with the model already inside, so power and marginal cost per inference drop to the minimum.

One goal,
three routes.

All three move the weights less, or not at all. They differ in where the weights live and whether the model can still change. Each production chip takes one route; for a custom build, the route follows how often your model changes.

17,000token/s

On-device decode speed

Forge design target · Qwen 3.5 0.8B

~1 W

System power

Forge, lightweight edge class

100%

Data stays on the device

Local inference, no network needed

≈ $0

Marginal cost per inference

After the one-time chip purchase

First-generation ASICs,
with the model already inside.

Pick by workload: a watt-class handheld, a node that runs 35B-parameter models, or a vision chip that keeps up with every frame. If none fits, we build one around your model.

Forge lightweight edge LLM inference chip with the Fanshi mark laser-etched on the package
Fig. 3 · Forge, lightweight edge chip

Forge

Edge LLM

A complete language model inside a watt-class handheld device.

17,000token/s
Core model
Qwen 3.5 0.8B
Decode speed
17,000 token/s
System power
Watt-class
Route
MaskROM compute-in-memory

Typical applications

  • Handheld translators
  • Voice-memo hardware
  • Offline in-car voice assistants
Forge datasheet

Figures on this page are current product definitions and design targets. Final specifications, test conditions, availability and delivery versions are confirmed in project documents agreed by both parties.

Model still evolving?
Don’t wait for tape-out.

License Fanshi inference cores into your own ASIC or FPGA. One core family, four architectures, running on an evaluation board first.

FS-Fusion Multimodal fusion inference IP, illustration
Fig. 7 · FS-Fusion

Multimodal fusion inference IP

FS-Fusion

Joint vision-language understanding (VLM) and multi-sensor fusion, computed in place at the edge.

Native SmolVLM + SegFormer MiT fusion pipeline

FS-Fusion details
FS-Attention Hybrid attention inference IP, illustration
Fig. 8 · FS-Attention

Hybrid attention inference IP

FS-Attention

Native support for language models built on hybrid attention, including the gated full-attention layers of the Qwen3.5 family.

Qwen3 / 3.5 (0.6B to 9B) and SmolLM2

FS-Attention details
FS-MoE Sparse mixture-of-experts inference IP, illustration
Fig. 9 · FS-MoE

Sparse mixture-of-experts inference IP

FS-MoE

Built for dynamically activated sparse models, carrying large parameter counts within an edge bandwidth budget.

Large sparse models on edge budgets

FS-MoE details
FS-Linear Linear attention and new-architecture IP, illustration
Fig. 10 · FS-Linear

Linear attention and new-architecture IP

FS-Linear

Linear attention and RNN-style architectures led by DeltaNet, including the Gated DeltaNet layers used from Qwen3.5 onwards.

Very low-latency streaming with a fixed-size state, however long the context

FS-Linear details

Run it on an FPGA first.
Decide on tape-out later.

The evaluation board and customisable IP licensing validate model mapping, interfaces and the full system path. An engineering sample exists before the silicon does, so design-in and tape-out decisions rest on measurements.

Ask about IP licensing
FPGA evaluation board on a lab bench connected to a laptop, with inference output scrolling in a terminal
Fig. 11 · FPGA evaluation board running an inference core

Any model can
travel this line.

That is how the first-generation chips were made, and yours can follow the same path: run it on an evaluation board first, then choose between an existing chip and one of your own.

  1. 01

    Pick the route

    Target model, performance and power goals, update cadence, interfaces. The first question is always: how often does the model change?

  2. 02

    Evaluate

    Run the model on the evaluation board, measure accuracy and throughput, and get a working engineering sample before any silicon is committed.

  3. 03

    Design in

    Interfaces, system software and field validation built around the evaluation results, turning the sample into a product prototype.

  4. 04

    Deliver the ASIC

    Once the model is frozen: a production chip, or the IP integrated into your own silicon.

Same tokens.
A 913× gap over three years.

Hardware is a one-off purchase. A token bill never stops. Pick the model you use today and drag the daily volume.

100K500M
$70$420

Self-hosted cloud GPU estimated at $0.70 per million tokens (2026 median H100 rental). After the one-time chip purchase, the marginal cost per inference is ≈ 0.

Three-year inference cost

Pays back within the first month

GPT-5.5 · API$192K
Self-hosted cloud GPU$7,665
Fanshi chip$210
API, cumulativeSelf-hosted GPU, cumulativeFanshi, cumulative
0107K215K0 mo12 mo24 mo36 mo

Prices are blended 2026 public API list prices (input:output ≈ 1:1). Illustrative only; actual costs depend on usage and quotation.