More intelligence. Less compute.

Four nodesbecome one.

Run Qwen3.8-2.4T on one NVIDIA B200 node instead of four for the BF16 baseline. Minima combines model-weight compression, KV-cache compression and an optimised runtime to get more from your infrastructure.

4× smaller: model weights
3.5× smaller: KV cache
2× higher: tokens per second

Benchmark your model Explore deployment options

See benchmark context

Flagship model Qwen3.8-2.4T

BF16 baseline

Node 01

Node 02

Node 03

Node 04

4 B200 nodes

Minima optimised

Node 01

1 B200 node

Minima internal benchmark · Qwen3.8-2.4T · NVIDIA B200 nodes · BF16 baseline. Results depend on the model, workload and serving configuration.

Internal benchmark

Qwen3.8-2.4T

NVIDIA B200 nodes · BF16 baseline

Four NVIDIA B200 nodes for the BF16 baseline become one NVIDIA B200 node with Minima.

4× smallerModel weights

3.5× smallerKV cache

2× higherTPS

Minima internal benchmark. Results depend on the model, workload and serving configuration. See benchmark context.

Latest internal testing

Qwen3.6-27B

1 × RTX PRO 6000 Blackwell

Minima weight and KV-cache optimisation preserves MMLU while delivering 2× faster tokens per second in the tested configuration.

3.2×Model compression3.5×KV-cache compression

2×TPS improvement

86.3 → 86.1MMLU · dense to Minima

Internal testing; paper in preparation. Configuration-specific result, not a general performance guarantee.

Latest from Minima

Research moving into production systems.

We work across the representation, memory path and serving runtime—because real inference efficiency is a systems problem.

01 / RESEARCH

A practical tensor-network pipeline at 32B scale

Our preprint covers learned sensitivity, mixed Tucker/TT/TR decompositions, healing, custom kernels and speculative decoding in a single production-shaped pipeline.

Explore the paper

02 / HOSTED API

Hosted inference

Minima-optimized inference through an OpenAI-compatible API

Access Minima-optimized inference through an OpenAI-compatible API, with pay-as-you-go access and a committed usage option.

Explore the API

03 / OPEN SOURCE

Private retrieval and agent workflows

Minima AI Agent connects enterprise data to source-linked search, chat and configurable workflows in customer-controlled environments.

Explore Minima AI Agent

The mnma engine

Optimize the entire inference path.

Minima is the company. mnma is our model-optimization and inference engine: a coordinated stack for weights, cache and runtime.

Model weights

Hardware-aware low-precision formats and tensor-network decompositions reduce weight memory while protecting sensitive regions.

KV cache

Tiered compression preserves recent and high-value context while storing older cache pages in compact representations.

Runtime

Custom CUDA kernels, scheduling and speculative decoding turn smaller representations into measurable serving gains.

The mnma inference path

01InputOpenAI-format request

02Optimized modelProtected + compact weights

03Compressed KVPrecision-aware cache tiers

04Custom runtimeKernels + scheduling

05APICompatible response

Benchmark evidence

Evidence, not benchmark theatre.

Every result is tied to a model, hardware configuration, workload and quality gate. Change the setup and the result can change.

Loading the published benchmark record…

Internal results

Qwen3.6-27B

3.2× weight compression, 3.5× KV-cache compression and 2× faster tokens per second on one RTX PRO 6000 Blackwell.

MMLU dense → Minima: 86.3 → 86.1
Quality change: −0.2

Internal results

Gemma 4-31B

2.9× model-weight compression and 3.5× KV-cache compression, with quality maintained in the tested configuration.

MMLU dense → Minima: 70.5 → 70.1
Quality change: −0.4

Products

Choose the deployment model that fits your data and economics.

Use the engine in your infrastructure, access optimized hosted models, or deploy a private application stack over enterprise data.

ENGINECore platform

mnma Engine

Optimize and serve your model inside your infrastructure.

Open-weight and internal models
Weight and KV-cache optimization
Custom serving runtime
On-premises or VPC

Benchmark your model

APIHosted inference

Minima Hosted API

Access Minima-optimized inference through an OpenAI-compatible API.

OpenAI-compatible request format
Minima-optimized inference
Pay-as-you-go access
Committed usage option

Explore the API

APPLICATIONOpen-source core

Minima AI Agent

Search, chat and build workflows over private enterprise data.

On-premises or VPC
Private model and data path
Source-linked retrieval
Enterprise deployment support

Explore the Agent

Minima research · arXiv:2602.01613

A practical tensor-network compression pipeline for production-scale LLMs.

Sergii Kozyrev · Davyd Maiboroda · Minima AI, Inc.

The preprint describes a five-stage system: learned sensitivity analysis, mixed structural decomposition, healing, custom CUDA kernel execution and speculative decoding. The central result is not a standalone compression ratio—it is an end-to-end serving result on a disclosed 32B workload.

Scope13 pages · 5 figures

SubjectsMachine Learning · AI

StatusarXiv preprint · Feb 2026

Explore Research View all papers

Results page from the Minima paper showing peak VRAM and throughput tables

First page of the Minima tensor-network compression paper

Who Minima is for

Teams treating inference as infrastructure.

We focus on workloads where memory, context capacity, throughput and data control have direct operating consequences.

ML platform teams

Fit and serve open-weight or internal models against explicit latency, memory and quality gates.

AI product teams

Increase serving headroom before product growth turns GPU capacity into a bottleneck.

Regulated enterprises

Keep model traffic within controlled environments and support existing compliance programs.

GPU-constrained operators

Explore practical routes to longer context, higher concurrency or fewer deployment nodes.

Company

Building the world’s most efficient intelligence factory.

Minima AI, Inc. builds the mnma engine, hosted inference access and private AI applications from one shared systems foundation. Inference optimisation is our enduring manufacturing advantage: it improves what we can serve today and compounds into the models we will build next.

Co-founder & CEO

Sergii Kozyrev

Former Principal AI Engineer at Atlassian. Sergii has built production search, recommendation and generative-AI systems and leads Minima's product and inference strategy.

Co-founder & CTO

Davyd Maiboroda

An AI systems architect based in California, Davyd leads Minima's model compression, kernel and runtime engineering.

Benchmark with us

See what Minima can do for your model and GPUs.

Send us the model, hardware and workload you care about. We will propose a benchmark with explicit quality, throughput, memory and cost gates.

Model or model family
GPU and deployment environment
Context, concurrency and traffic shape
Your quality and latency constraints

Request a benchmark

Submit the form and Minima will contact you about your request. See our Privacy Policy and Terms of Service.