We are developing an experimental prototype and preparing hardware testing. Four research layouts are built and checked in an open toolchain; the compact neural core τ is at the concept stage.
Next step: hardware testing of the prototype, then measurement of power and timing.
Concept visualization
Φ
Ternara Φ
What it explores
A reference answer for checking start-up and the digital interface.
Where it can help
Debugging electronics, test benches, training.
Stage
Layout · simulation in CI
Concept visualization
e
Ternara e
What it explores
Experimental matrix compute for future analysis of sensor data.
Where it can help
After further work: telling normal from unusual states of a fan, compressor or pump.
Stage
Research layout · simulation in CI
Concept visualization
γ
Ternara γ
What it explores
Event-driven neural compute for researching equipment signals.
Where it can help
After further work: unusual vibration patterns, repeating pulses, mode changes.
Stage
Research layout · simulation in CI
Concept visualization
C
Ternara Corona
What it explores
Reference number formats for checking compute systems.
Where it can help
Engineering tools: testing decoders, cross-checking software and hardware results.
Stage
Research layout; manufacturing of this version is not planned
Concept visualization
+
Ternara τ
What it explores
We are developing a compact core for local neural processing in home devices.
Neural networks multiply a great many numbers. Ternara cores use ternary weights: every weight is only −1, 0 or +1. Multiplying by such a weight means subtract, skip or add, so the arithmetic needs no large multipliers and the weights take much less memory.
Less memoryA ternary weight is stored in 2 bits instead of the usual 16, so the same memory holds 8 times more weights.
Simpler arithmeticAdd, subtract or skip instead of a full multiplication.
From a written specificationEach block starts as a specification. The hardware description is generated from it and checked against reference models.
Pre-order
Pre-order Ternara chips.
Order single samples right here: pick a design, set the quantity, add to cart and check out with your delivery address. No payment now: we confirm by email and send an invoice when the batch is confirmed. Batches get a lower per-chip amount on request.
By quantity
Sample1–10 pcs€600 per evaluation kit
+ €30 shipping per kit. Order on the site.
Small batchfrom 100 pcs≈ €270 per chip
Packaged QFN parts from a dedicated run.
Batchfrom 1,000 pcsfrom ≈ €8 per chip
Dies on a mini-board from a dedicated run.
Seriesfrom 10,000 pcsOn request
A quote for your volume and package.
The low per-chip amount is for batches only: each batch is made in a dedicated production run for your order. Batch amounts are indicative; the final quote, lead time and delivery format come after your request. Single samples are sold at the amount shown on the site.
Designs
Concept visualization
Φ
Ternara Φ
Pre-order
A reference answer for checking start-up and the digital interface.
Ternara edition in development · estimated shipping 2027
Sample:€600 per evaluation kit
Concept visualization
e
Ternara e
Pre-order
Experimental matrix compute for future analysis of sensor data.
Ternara edition in development · estimated shipping 2027
Sample:€600 per evaluation kit
Concept visualization
γ
Ternara γ
Pre-order
Event-driven neural compute for researching equipment signals.
Ternara edition in development · estimated shipping 2027
Sample:€600 per evaluation kit
Concept visualization
C
Ternara Corona
Waitlist
Reference number formats for checking compute systems.
Pre-orders open once the design is in a production run.
Concept visualization
+
Ternara τ
Waitlist
We are developing a compact core for local neural processing in home devices.
Pre-orders open once the design is in a production run.
✓ Pre-order received
✓ Request received
✓ You are on the list
Terms
Samples: no payment now. We confirm every pre-order by email and send an invoice when the batch is confirmed.
You can cancel free of charge at any time before paying.
Estimated shipping: 2027. Dates are estimates and can move; we will keep you informed.
A sample is a prototype for development and evaluation, not a finished consumer device.
Batches: amounts are indicative; the final quote, lead time and delivery format are agreed after your request.
Taxes and import duties, if any, are shown on the invoice. We use your details only to process your request.
Green AI
Green AI, by arithmetic.
Most of the energy in AI goes into moving weights and multiplying them. Ternary weights attack both: three values instead of thousands, and no multiplier at all. Here is the evidence, with sources.
× → ±
No multiplier
With weights limited to −1, 0 and +1, each multiplication becomes subtract, skip or add. Microsoft's BitNet paper describes its matrix multiplication as needing almost no multiplications. Source: arxiv.org ↗
≈10×
Less weight memory than BF16
A ternary weight carries log2(3) ≈ 1.585 bits, against 16 bits for BF16: about 10× less memory for those weights. Practical llama.cpp packings use 1.6875 or 2.0625 bits per weight. Source: github.com ↗
55–82%
Less CPU energy per token, reported by Microsoft
Microsoft reports that its bitnet.cpp software, running ternary BitNet b1.58 test models, used 55.4–70.0% less energy per token than llama.cpp on an Apple M2 Ultra CPU. On an Intel Core i7-13700H laptop CPU, Microsoft measured 71.9–82.2% less energy per token with bitnet.cpp than with llama.cpp (700M and 7B test models). Source: arxiv.org ↗
1.3–2.6 nJ
One DRAM read, versus 0.03 pJ for an 8-bit add
The same table puts a 64-bit DRAM read at 1.3–2.6 nJ, the cost of tens of thousands of 8-bit additions. Moving fewer bits per weight matters more than the arithmetic. Source: gwern.net ↗
945 TWh
Data centres by 2030, according to the IEA
Data centres used about 415 TWh of electricity in 2024, around 1.5% of the world's total, and are set to reach around 945 TWh by 2030, slightly more than Japan uses today. Source: iea.org ↗
As of 2026-10-10: no die has been returned from fabrication. No Ternara silicon has been measured. Power has never been measured on any Ternara hardware. Every energy number on this page belongs to other teams' software, models or chips. None of them describes Ternara. Microsoft's CPU energy figures compare bitnet.cpp with llama.cpp on test ("dummy") models, not trained ones. The 71.4× arithmetic saving and the 0.028 J per token are model estimates, not measurements.
Vision · products
Where Ternara chips will work.
From an air conditioner to servers: the scenarios we are building toward. These are development goals, not finished products.
Scenario · air conditioner
Say it. The room follows.
“Set 22 degrees and switch on quiet mode.”
The command stays at home. In the target architecture the device works out what you want on its own and hands a permitted command to the air-conditioner controller.
The path of one command
YouA short command in your own words.
Speech to textA speech recognizer in the device. A separate component.
Ternara coreChooses one of the permitted commands: temperature, mode, fan. This is the part we are developing.
Device controllerChecks the command against its own limits and protections, then applies it.
ResultThe room cools to 22 degrees in quiet mode.
Interface demo
Try the interface
Choose a command:
Target24°
ModeCool
Fan—
Timer—
Permitted command —
Interface demo: a software model (JavaScript), not a chip output. It does not control a real air conditioner.
Works without internetControl stays available when the connection drops.
Data stays homeCommands are processed inside the device, not on a remote server.
Simple to usePlain words instead of menus and remote-control codes.
Energy savings, response time and free-form speech are development goals. We will state them as results only after measurement.
Beyond air conditioning
ClimateHeaters, fans and ventilation: “Warmer in the bedroom, quieter at night.”
Lighting“Dim the living room for the evening.”
Home sensors“Tell me if the humidity in the bathroom stays high.”
Future direction
Ternara τ: AI compute for computers, games and servers.
We plan to develop the Ternara τ compute core into a specialised accelerator for compatible neural network models. The prospect is local AI features on computers and game systems, and neural network inference on servers.
The goal is efficient execution of selected models with ternary weights. The architecture, software support and benefits are to be confirmed on prototypes.
ComputersPossible acceleration of compatible models for a local assistant, document search and speech processing.
Game consolesPossible acceleration of suitable models for speech, character dialogue and other neural features, in integration with the platform maker and game developers.
ServersInference for compatible neural network models, assessed by throughput, latency and energy of the whole system.
How the direction develops
Run a selected trained model end to end and check the quality of its results.
Build a prototype with memory and all required operations; measure speed and energy.
Connection interface, driver and runtime for the target platform.
Scale up where the benefit is confirmed, and test in a partner's real scenario.
The same local processing can serve equipment far from a reliable connection. Each direction starts with a pilot on real data.
Pumps and water supplyLocal analysis of vibration, current or other available signals; spotting deviations in operation.
Pilot direction
AgricultureProcessing sensor readings for irrigation, greenhouses and remote sites.
Pilot direction
IndustryMonitoring equipment condition and classifying events on site.
Pilot direction
AviationResearching local processing of diagnostic sensor data.
Research direction
Space systemsResearching compact compute for autonomous telemetry processing.
Research direction
Vision
Mesh networks
Towers between cities. Mesh inside them.
Today every block needs a cell tower nearby. In our vision, towers carry traffic between cities, while inside a city devices pass packets to each other: phones, routers, appliances. The more devices, the denser the network.
Concept visualizationConcept visualization
Already working today
Meshtastic turns inexpensive LoRa radios into an off-grid text-messaging mesh: each radio rebroadcasts what it hears, at 0.33 to 21.88 kbit/s depending on the preset. Source: meshtastic.org ↗
NYC Mesh links buildings with rooftop routers so traffic passes from building to building across New York, and reaches the internet by peering at an internet exchange point, with fibre-connected supernodes. Source: wiki.nycmesh.net ↗
guifi.net, a community network mostly in Catalonia and Valencia, reported over 37,000 active nodes and about 71,000 km of wireless links in December 2021. Source: en.wikipedia.org ↗
Freifunk in Germany counts about 400 local communities and over 41,000 access points running OpenWrt-based mesh firmware. Source: en.wikipedia.org ↗
What it would take
Gupta and Kumar (2000) showed that in a multi-hop wireless network of n random nodes, each node's throughput falls as W/√(n log n): the bigger the mesh, the less each node gets.
In MIT's 2001 simulations, a long chain of 802.11 relays delivered about 1/7 of the single-hop rate; capacity scales with network size when traffic stays local.
NYC Mesh describes mobile-phone mesh networks as very low bandwidth, not real-time, and needing phones close together, and connects buildings instead.
Working city meshes depend on backhaul: NYC Mesh peers at an internet exchange and uses fibre, and guifi.net uses long-distance Wi-Fi links and community-owned fibre.
LoRa meshes such as Meshtastic run at 0.33–21.88 kbit/s: enough for text, not for video or live calls.
As of a 2023 survey, 5G sidelink had no standardized multi-hop routing; phone-to-phone relaying across a city is not part of today's cellular standards.
Radio is regulated: in the US, unlicensed devices have no vested right to a frequency, must not cause harmful interference and must accept interference (47 CFR §15.5); in the EU, licence-free LoRa bands are limited by duty cycle, and cellular bands need licences.
Ternara blocks have no radio today. A radio chip of our own is possible as a separate future product, but it is analog RF design, a much harder field than digital blocks. Radio certification (Anatel, FCC, CE) is required.
Pilot
Let's build your device scenario together.
A pilot is a short joint project around one device scenario, for example climate control. We agree on the goal, the data and the success criteria, and show the result in an environment named up front: a software model, a simulation or a hardware bench.
Choose one scenario and agree which part Ternara does and which part your components and device do.
Define the input data, the permitted commands and the success criteria.
Prepare a demonstrator and a test report in the environment that is actually available.
Next comes a short call to choose the scenario and the success criteria. Supply terms are discussed after the pilot results.
Company
Ternara: intelligence where there is no connection.
We are developing compute cores for compact devices that process sensor data locally. Our goal is to help monitor equipment, agriculture and remote sites of any kind with Ternara AI, without a constant connection to the cloud.
For investors
Nearest verifiable result
Hardware testing of the prototype with measured power and timing.
Possible path to revenue
Paid pilots with device makers, then a joint prototype in a partner's device, then licensing of the cores or supply of chips under an agreement.
Concept visualization
Concept visualization
Concept visualization
Concept visualization
Concept visualization
