Articles grouped by business problem, industry, and system type.
A private AI assistant for payroll gathers the right payslip and the internal decision record for an employee's question, and drafts a reply. The reply describes what follows from the documents, and a payroll specialist checks and sends it. At Syntalith an application on a local model costs from €3,000 net plus GPU hardware. At 400 payroll questions a month of 15 minutes each, it pays back in 1.2 months.
A private AI assistant finds earlier material trials in the R&D archive and shows each with its material version and test conditions. The engineer sees that the high-temperature trial covered an earlier version, while the changed version has only a room-temperature trial. At Syntalith an application on a local model costs from €3,000 net plus GPU hardware. At 200 questions about earlier tests a month of 30 minutes each, it pays back in 1.2 months.
A local AI assistant helps a branch employee understand a customer's question in another language and put together a short explanation, even without internet. Before you buy, settle whether this help is what is missing most, or whether the branch needs a backup connection to company systems. At Syntalith an app on a local model costs from €3,000 net plus GPU hardware. With 600 conversations in another language a month at 4 minutes each, it pays back in 3.6 months.
A small language model on a warehouse terminal turns a worker's few words, for example about a damaged package, into a clear description of the event, and the worker approves the text before it is saved. It only makes sense when the model runs on the terminal you use and adds no facts the worker did not see.
A small local model can catch a conflict between a form field and a customer's note, for example collection in person chosen and a request for a courier, and prepare a question for the customer. First check whether the form invites such a conflict. At Syntalith an app on a local model costs from €3,000 net plus GPU hardware. With 1,500 form orders a month at 4 minutes each, it pays back in 1.2 months.
A small local model can suggest to the helpdesk duty agent which team should get a ticket. You assess at once the accuracy of the suggestions and how long a case waits for assignment in the morning, on the hardware the company has. At Syntalith an app on a local model costs from €3,000 net plus GPU hardware. With 2,000 tickets a month at 2 minutes each, it pays back in 2.0 months.
A small local catalogue search lets a salesperson with no network access find product variants from the client's description and show the differences. It works when the product data and the material to check are on the device. First check how much names, filters and aliases solve in ordinary full-text search.
Before you buy a local document assistant, compare its answers to your own questions. Does it point to the right version of the source, does it admit when information is missing, and how long does checking take? At Syntalith the app costs from €3,000 net. With 500 questions a month and a GPU server at €214, it pays back in 2.0 months.
A local AI model suits confidential proposals when data rules rule out your current tool. The decision covers quality on your documents, user access, performance and responsibility for maintenance, and buying a server is only one of those elements. At Syntalith an app on a local model costs from €3,000 net plus GPU hardware. With 40 proposals a month at 120 minutes each, it pays back in 1.7 months.
Local AI in a law firm collects dates from the case files into a chronology, describes what they concern and keeps a link to the specific page. It marks an illegible date as unconfirmed, and a lawyer judges its legal meaning. An application on a local model costs from €3,000 net plus GPU hardware at Syntalith. At 150 case chronologies a month and 30 minutes each, it pays back in 1.7 months.
A workstation suits local AI when one person plans their own work and accepts breaks. You consider a shared server when the team needs access that does not depend on one employee's computer. The choice needs a test of concurrent tasks and answer quality, and a decision on who maintains the service.
A small local model recognises the type of a document, for example an invoice or a delivery confirmation, splits a mixed PDF into parts and routes each to the right queue. It stops before posting and tax assessment. Uncertain pages go to a person. With 3,000 documents a month at a minute and a half each, a €3,000 app with a GPU server pays back in 1.4 months.
OpenCode has a documented path to local models through an OpenAI-compatible interface. Claude Code does not officially support models other than Claude, so in our test it needed an emulation of the Anthropic interface. On the same long task with a local Qwen3.8-27B, Qwen Code scored 100 out of 100 in 502 seconds, and Claude Code through emulation 98 points in 1,732 seconds.
Codex connects to a custom model provider through the Responses API, so it also works with a local Qwen3.8-27B on vLLM. In our experiment on an RTX 3090 the connection worked, but the code fix from Codex scored 87/100, while the one from Qwen Code on the same server scored 100/100. A working connection says nothing about the quality of the changes.
A Polish knowledge assistant with RAG finds the right document, checks its version and the employee's permissions, and then answers with a quotation or refuses. You choose the model on the full task: in 2026 Bielik-11B led in RAG on technical documentation, and PLLuM-12B matched 70B models on Polish Wikipedia while using 83% less energy.
Fixing the same Go bug took a local Qwen3.8-27B 502.14 s on a modified vLLM and 1468.76 s on llama.cpp, with scores of 100/100 and 98/100. The whole profile differed: the engine, the weight format, the KV memory and the context limit, so the result points to a configuration for further trials and does not measure the advantage of the engine alone.
After several failed attempts, we kept Qwen Code with medium reasoning effort set in the client and in the request, preserved thinking, a 150,000-token context and no fixed limit on response length. This configuration comes from one RTX 3090 card and one active agent, with a short profile test after every change.
The Qwen3.8-27B profile we tested on an RTX 3090 with 24 GB uses W4A16 AutoRound weights, an FP8 KV cache, MTP-3, a modified vLLM and a context of 150,000 tokens with one active request. After the start you check the process, the model and the limit at the API address, plus a short reply, before you run the agent.
In our test, DFlash2-7 speculative decoding gave a median of 120.9 tokens a second against 93.3 for the MTP-3 profile, which is 29.6% more. It ran with a context of 65,536 tokens in place of 150,000, so it stayed an optional profile for short tasks, and long tasks stayed on MTP-3.
The same local model, Qwen3.8-27B, fixed code differently depending on the program that drove it. In a long Go task, Qwen Code finished in 8:22 with a score of 100/100, Codex in 19:10 with a bug that made us reject the change, and Claude Code in 28:52 with 98/100. In a shorter CSV import, OpenCode prepared the fastest correct change, in 89 seconds.
Private models
We select the hardware and model, deploy it and measure quality on your examples. You can start with a computer you already own.
Times are shown in your own time zone. We work with clients across time zones.
See private modelsOpenAI makes ChatGPT and Anthropic makes Claude. We belong to the partnership programs of both companies, so we work with their tools every day and learn from the creators themselves.
Claude Partner NetworkParticipation in programs does not mean that OpenAI or Anthropic endorses our services.