AI inference research lab
Minima makes intelligence radically more efficient. We compress model weights and KV cache, then serve them with a hardware-aware runtime to increase useful intelligence per dollar—without replacing your model or rewriting your application.
Run more AI on every GPU across hosted APIs, private infrastructure and edge deployments.
QWEN3-32B / A100 80 GB
New Optimised ModelsUp to 3.2× weight compression · 3.5× KV cacheQwen3.6-27B · Gemma 4-31B
Published system result
Qwen3-32B · one NVIDIA A100 80 GB · bf16 · 8K context
Peak VRAM falls from 64 to 40 GiB while MMLU remains close to the dense baseline.
Loading paper-backed metrics…
Figures reported in the Minima arXiv preprint submitted February 2026. Results vary by model, hardware, context and traffic shape. See the disclosed setup.
Latest internal testing
1 × RTX PRO 6000 Blackwell
Minima weight and KV-cache optimisation preserves MMLU while delivering 2× faster tokens per second in the tested configuration.
Internal testing; paper in preparation. Configuration-specific result, not a general performance guarantee.
Latest from Minima
We work across the representation, memory path and serving runtime—because real inference efficiency is a systems problem.
Our preprint covers learned sensitivity, mixed Tucker/TT/TR decompositions, healing, custom kernels and speculative decoding in a single production-shaped pipeline.
Explore the paperWe are preparing hosted access for optimized open-weight models. OpenRouter provider integration is in progress.
Join launch accessMinima AI Agent connects enterprise data to source-linked search, chat and configurable workflows in customer-controlled environments.
Explore Minima AI AgentThe mnma engine
Minima is the company. mnma is our model-optimization and inference engine: a coordinated stack for weights, cache and runtime.
Hardware-aware low-precision formats and tensor-network decompositions reduce weight memory while protecting sensitive regions.
Tiered compression preserves recent and high-value context while storing older cache pages in compact representations.
Custom CUDA kernels, scheduling and speculative decoding turn smaller representations into measurable serving gains.
Benchmark evidence
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
3.2× weight compression, 3.5× KV-cache compression and 2× faster tokens per second on one RTX PRO 6000 Blackwell.
Internal results
2.9× model-weight compression and 3.5× KV-cache compression, with quality maintained in the tested configuration.
Products
Use the engine in your infrastructure, access optimized hosted models, or deploy a private application stack over enterprise data.
Optimize and serve your model inside your infrastructure.
Use Minima-optimized models through an OpenAI-compatible endpoint.
Search, chat and build workflows over private enterprise data.
Minima research · arXiv:2602.01613
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.
Who Minima is for
We focus on workloads where memory, context capacity, throughput and data control have direct operating consequences.
Fit and serve open-weight or internal models against explicit latency, memory and quality gates.
Increase serving headroom before product growth turns GPU capacity into a bottleneck.
Keep model traffic within controlled environments and support existing compliance programs.
Explore practical routes to longer context, higher concurrency or fewer deployment nodes.
MINIMA BLOG
Article
This is a joint benchmark of Qdrant retrieval and Minima-optimized inference with Qwen3.6-27B model on one RTX PRO 6000 Blackwell.
Read articleArticle
Minima solved the efficiency problem inside the inference worker: 53% more tokens from Qwen3.6-27B on Blackwell. Rafay Token Factory solved the operability problem around it: a governed, token-metered service.
Read articleCompany
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
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
An AI systems architect based in California, Davyd leads Minima's model compression, kernel and runtime engineering.
Benchmark with us
Send us the model, hardware and workload you care about. We will propose a benchmark with explicit quality, throughput, memory and cost gates.
Submit the form and Minima will contact you about your request. See our Privacy Policy and Terms of Service.