| Result | two models fit into 22.4 GB of 24 GB of video memory with a 65,536-token context |
| Source | Digital Platforms, case no. 4084, 2 September 2026 |
| A similar task for us | AI employee |
| Industry | IT companies |
What was done and what came out
- The agent edits texts in set personas, runs a night queue of research tasks started by a scheduler every 10 minutes, searches the web and works with files inside its working folder only.
- It is reliable for a chain of about 6 tool callsfunction calling. Function calling is a way for a language model to ask a program to do an action: the model picks a function and fills in its parameters, and your application code runs the request and returns the result. in a row; the timeout for a cold model load had to go above 300 seconds; fixing twelve errors took one evening.
- It serves one user at a time.
- Published on 21 August 2026.
What to take for yourself
Exact memory math from the case: model weightsmodel parameters. Model parameters are the numbers inside a neural network that it sets during training and then uses to compute every answer. plus the KV cache must fit into video memory with a margin. Our memory calculator uses the same formula.
AI agents on a local model for your team: a pilot in 14 days. A pilot AI agent on a local model in 14 days from $1,499 with the metric in the contract, on your hardware or on rented GPUs. An AI department of 3–5 agents in 4–6 weeks from $4,999. You buy the hardware to our specification.
Sources and fact-checking
The facts and numbers on this page come from the publication. Data type: project catalogue entry. The company is not our client: the page shows what a smaller company can repeat from this experience. Checked by the editors on 30 September 2026.
- Digital Platforms, case no. 4084 (platforms.su)· 2 September 2026
Found a mistake? Tell us
Similar cases
- AI department B2B services case: from one lead agent to a team that protects every lead
- B2B services lead handling: answer every request before the client leaves
- LANIT: requirements gathering 4 times faster, document search time 66% lower on a local LLM platform
- Samsung Electronics: ban on outside generative AI services after a source code leak, own tools in development
- ETH Zurich, EPFL and CSCS, Switzerland: fully open sovereign model Apertus: 8 and 70 billion parameters, 15 trillion tokens, more than 1,000 languages
- Petrovich-Tech: a local model with 9 billion parameters matched cloud models on tasks of 2–3 steps: 27 runs
- Developer MR: Project finance AI controller in an isolated environment: 80% less time on credit document analysis
- All materials in the industry: IT
Prices and measurements in this section were checked against the sources at the bottom of each page. We build AI agents on cloud and local models and do not sell hardware. Calculations are estimates; legal conclusions on data protection come from your lawyer.