Inceptron

Inceptron operates a platform for running AI models in production, combining serverless inference endpoints, dedicated GPU deployments and a proprietary optimization compiler. Users can call open-source or fine-tuned models through an OpenAI-compatible API at api.inceptron.io, import custom checkpoints, or start from a curated model library, then create versioned endpoints with API keys, access controls and rollout policies. The hosted catalogue includes large open-weight models such as GLM-5.2, Kimi-K2.7 Code, Kimi-K2.6, DeepSeek-V4 Flash and MiniMax-M2.5, listed with per-token pricing, quantization format, parameter size, context length and throughput figures, alongside Llama and Qwen family models.

The optimization layer is built around a compiler that fuses computation graphs, auto-tunes kernels and plans memory for a target hardware configuration, plus performance-aware compression through quantization and pruning. Company material describes automated passes including shift-based reparameterization, mixed-precision optimization, memory and cache tuning, and workload-specific search, applied across GPUs, CPUs and accelerators such as AWS Inferentia and FPGAs, for training, diffusion-model deployment and edge or cloud inference. Operational features include dynamic request batching, autoscaling, scheduled inference, integrated logging and tracing, and connections to cloud storage buckets, CI/CD and external observability stacks.

The platform is positioned for enterprise use with team controls, encryption and hardened isolation, ISO 27001 and GDPR compliance claims, data-residency controls including EU-only residency, SLA-backed uptime and zero-retention data handling.

Founding story

The founding team came from machine-learning research backgrounds, with experience described as spanning AI labs and infrastructure companies, and built an optimization compiler aimed at making model deployment across varied hardware and frameworks more efficient [5][7]. Named co-founders include Steffen Malkowsky, listed as CTO and founder, and a co-founder identified as Anton M. [5][7].

Business model

Inceptron sells access to its inference platform directly to engineering and machine-learning teams, with self-serve serverless usage alongside contracted dedicated deployments and enterprise add-ons such as custom regions, private networking and premium support quoted separately [0][4].

Serverless inference is billed per token, with separate input and output token rates rounded to the nearest 1,000 tokens per request batch (for example Llama-3.3-70B Instruct at $0.10 per million input and $0.30 per million output tokens); dedicated deployments are billed hourly per GPU pro-rated to the minute (for example an H100 80 GB instance at $3/hour). Commitment discounts are offered at 5% for one month, 10% for six months and 20% for twelve months, applying to both dedicated and serverless usage, with overage charged at on-demand rates and unused prepaid amounts rolling over within the commitment window [4]. Published catalogue prices range from $0.13 per million input tokens for DeepSeek-V4 Flash to $1.20 for GLM-5.2 [3].

Traction

Reported headcount is in the 11-20 range, with employees across engineering, design and sales functions [7]. The compiler is in an early-access phase and the hosted platform serves a catalogue of open-weight models with published pricing and throughput figures [0][3].

Latest developments

The compiler has been opened for early access with automated model compilation, and new models including DeepSeek V4 Flash have been added to the serving catalogue [0][3].