Articles grouped by business problem, industry, and system type.
Qwen Code with local Qwen3.8-27B fixed the CSV export and the pagination with every hidden test passing, but in the user import it let through a record with an empty name or role. All 12 public tests passed, and the hidden ones gave 4 of 5. Codex and OpenCode passed 5 of 5 on this task.
Local Qwen3.8-27B, running through Qwen Code, built a recruitment dashboard from one sentence in 26 min 3 s. The first version, however, rendered 251,327 DOM elements and a page 485,065 px tall. The accepted version came after 48 min 25 s of model work across three sessions, and after the repair the page had 722 DOM nodes.
Qwen3.8-27B found the three facts asked for in both a 50,059-token and a 230,085-token input, and the shorter run took 59.3 s while the longer took 563.4 s. You size the context window to the 95th percentile of real input length plus a margin, because a bigger window costs time without a better answer.
Quantising Qwen3.8-27B to Q4, Q5 or W4A16 lets the model fit on a 24 GB card, but the memory you win goes on context. You choose the smallest format that passes the test on your own tasks, because our comparison changed the format, the engine and the context limit together.
One RTX 3090 with 24 GB of VRAM was enough to run Qwen3.8-27B and fix a bug in Go code in 502.14 s, with 52 tool calls, about 33.88 Wh of energy and a score of 100/100 on our criteria. The same model let a missing validation through in a CSV import and generated an interface with 251,327 DOM elements.
Qwen3.8-27B suits code, images and long inputs, Bielik-11B-v3.0-Instruct suits Polish text with no personal data, and PLLuM-12B-chat-2512 suits dialogue and RAG in Polish. A test on the same Polish examples decides, because 2026 studies give a different order depending on the corpus.
A Qwen model on your own server does not keep data in by itself: a document can leave through text extraction, logs or an agent's tool. A safe deployment starts with a map of the data flow, separate user identities and limits on tool calls. At Syntalith an app on a local model costs from €3,000 net plus GPU hardware. At 400 document tasks a month of 12 minutes each, it pays back in 1.6 months.
A local Qwen can be reached remotely in three ways: a web interface on a private network, a private network for apps and clients, or an SSH tunnel for one engineer. Each needs an explicit secret, a description of the data flow and tests before users get access.
Do not deploy a local Qwen when, in the pilot, a task does not finish in reasonable time, a small window turns the work into an hours-long session, the maximum window raises cost without better quality, a hidden check finds an error after green tests, or the interface fails a browser check.
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.