BTB Intelligence — plain answers.
The company, the inference system, the results, the people and the facts — stated simply.
What is BTB Intelligence?
BTB Intelligence is an inference-optimization company from Latvia (EU), founded in 2026 (SIA “BTB Intelligence”, registration number 40203724476). Its system of software layers — BTB Micro, BTB CASCADE, the QUANT family and a proof layer — makes large language models cheaper to serve on NVIDIA GPUs. As a second line of work it develops its own models, BTB Blitz and BTB Tsar, which are in development. It is a member of the MLCommons® Association and of the NVIDIA Inception Program.
What does the BTB inference system do?
It is a stack of independent layers, each removing a different source of serving cost: BTB Micro (input efficiency: more users on the same GPUs), BTB CASCADE (a right-sized model for every request; in development), a regime dispatcher (the right combination of layers for interactive or batch traffic; in development), a precision engine with the QUANT family (full-quality low-precision inference) and fewer sequential generation steps on structured tasks. A proof layer runs alongside all of them so that every result is verifiable and tamper-evident. How the layers work is not published; a patent application is in preparation.
What results has BTB Intelligence measured?
Internal measurements on NVIDIA B200, September 2026, on an industry-standard inference benchmark: Mixtral 8x7B on an 8× B200 system, 198,896 tokens per second (+54.4% over the best published 8× B200 result, Open-class quality); Llama 2 70B on an 8× B200 system, 107,384 tokens per second (+5.8% over NVIDIA's published 8× B200 result) with official accuracy targets met; Mixtral 8x7B with the QUANT family alone, 193,233 tokens per second (+50.0%); Llama 2 70B on one B200, 13,511 tokens per second (×3.23 against standard 16-bit serving, accuracy targets met) and 13,488 tokens per second for the Closed-style version (+6.3% per GPU); Mixtral 8x7B on one B200, 24,254 tokens per second (×4.18, Open-class quality) at 0.0275 J per token; BTB Micro up to 2× more concurrent users per GPU. Result not verified by MLCommons Association.
Are these official benchmark results?
No. They are BTB Intelligence's internal measurements and have not been submitted for official review. Result not verified by MLCommons Association. These results have not been through MLCommons review and may use measurement methodologies and/or workload implementations that differ from the specification for verified results.
What is BTB Micro?
BTB Micro is the input-efficiency layer of the BTB inference system for organisations that serve AI on their own GPUs. Measured by BTB Intelligence on NVIDIA B200 against a tuned vLLM setup across five open models, the same GPU serves up to 2× more concurrent users at the same latency target (median 1.9–2.0×, best case 2.37×), with answer quality held at parity in blinded judging. Once deployed, it runs inside the customer's own infrastructure. It is available for evaluation.
What is the QUANT family?
The QUANT family is BTB Intelligence's proprietary low-precision inference method: large language models run at reduced numerical precision while meeting official accuracy targets. In internal measurements it delivered 13,511 tokens per second on Llama 2 70B on one NVIDIA B200 — ×3.23 against standard 16-bit serving, with accuracy targets met — and 107,384 tokens per second on an 8× B200 system. A patent application is in preparation.
What is BTB CASCADE?
BTB CASCADE is the routing layer of the BTB inference system: it gives every request a model of the size it actually needs, so routine traffic does not pay for the largest model. It is in development.
What is the proof layer?
The proof layer runs alongside every other layer. Every run is written to a sealed, append-only record that shows any later change, and quality is certified against the baseline before a speed-up counts, so every figure can be re-checked.
Who founded BTB Intelligence?
BTB Intelligence was founded by Artjom Bodnarsky (Co-Founder & CEO), a systems architect who leads the company's technology and inference-optimization work and is the author of BTB Micro, and Mikhail Bickovsky (Co-Founder & CFO), a finance executive with more than 20 years in corporate finance, taxation, due diligence and M&A support.
Is BTB Intelligence a member of MLCommons?
Yes. Since 5 October 2026 BTB Intelligence has been a member of the MLCommons® Association and takes part in its Inference, Benchmark Infra and Power working groups.
Is BTB Intelligence part of the NVIDIA Inception Program?
Yes. BTB Intelligence is a member of the NVIDIA Inception Program, NVIDIA's program for AI startups.
Does BTB Intelligence have patents?
A patent application for our core methods is in preparation; until it is filed we publish results, not methods.
What are BTB Blitz and BTB Tsar?
BTB Intelligence's own language models, in development: BTB Blitz is a fast model trained verifier-first to answer directly and act through agents and tools; BTB Tsar is a deep-reasoning model for long, multi-step problems. Neither is released yet.
What is Optimum?
Optimum is BTB Intelligence's decision technology, in development. It turns a real situation into a formal decision problem, solves it, and returns the best available answer together with the reason it is the best.
What is BTB Brain?
BTB Brain is the app through which BTB Intelligence will deliver its models to users. It is in private testing and not publicly available.
Is BTB Intelligence related to Bit Brother Limited (former Nasdaq: BTB)?
No. BTB Intelligence is an independent AI company registered in Latvia in 2026. It has no relation to Bit Brother Limited, the former Nasdaq-listed company that once traded under the ticker BTB.
Where is BTB Intelligence registered?
BTB Intelligence is registered in Latvia, in the European Union, as SIA “BTB Intelligence”, company registration number 40203724476. btbintelligence.com is the official website of SIA “BTB Intelligence”.