Blog · Published · Updated
Qwen, DeepSeek, Kimi and GLM: what to self-host?
Short answer
Open-weight models from 2026, such as Qwen, DeepSeek, Kimi, GLM, Gemma and Mistral, differ in size, the hardware they need and how they are run. You start the review by asking where the system will run and what work it has to do, and you choose by testing on your own tasks.
How we workDetails ↓
- 01Call30 minutes with an engineer, free
- 02PrototypeA working system on your process and data, free
- 03ImplementationTarget and fixed price in the contract; 50% at signing
- 04Acceptance50% once the agreed target is met
- model on your server or in the cloud in the EU
- quoted after the review
- same work in application with EU cloud model
- from €3,000
- task, data and equipment analysis
- €0
We belong to the OpenAI and Anthropic partner programmes and learn through their official courses. What this means ↓
Author: Syntalith · 4 min read
Statistics
Open-weight models: statistics 2026
1.6 mo.
payback of a €3,000 local app at 400 documents a month: €2,792 of work today, €912 with checking and a GPU server. Scenario.
40%
According to Menlo Ventures (2025), Anthropic holds 40% of large companies' spending on language model APIs.
Source: Menlo Ventures, State of Generative AI in the Enterprise (2025) ↗
14–56%
According to Anthropic (2025), AI cut task time by 14% to 56% in studies with a control group.
Source: Anthropic, Estimating AI productivity gains (2025) ↗
11 of 13
According to the NoLiMa study (2025), at 32,000 tokens of context 11 of 13 models fall below half of the accuracy they reach with a short context.
20.0%
According to Eurostat (2025), 20.0% of EU companies with at least 10 employees used AI.
Figures checked: 22 October 2026
An employee's computer
smaller models, for example Qwen or Gemma
a workstation with a graphics card
The company's server
mid-sized models
a GPU server from €214 a month
A GPU cluster
the largest models, for example Kimi or DeepSeek
priced after the review
What the new open-weight families offer
| Family and version | What the official source says | What it means for a rollout |
|---|---|---|
| Qwen3.8-27B | A model with available weights; a smaller variant of the Qwen3.8 family | A candidate for a controlled installation on one card after quantisation |
| Qwen3.8-Flash-Next | 125B parameters in the main part, 6B active, plus n-gram embeddings and MTP | 6B active does not mean the file size or the memory needed for the whole model |
| DeepSeek-V4.1-Flash | A multimodal MoE, 552B parameters in the main part, MIT licence | Assess the full model and whether the server supports its architecture |
| Kimi K3 | 2.8T parameters, 104B active, its own Kimi K3 licence | A rollout needs infrastructure suited to very large weights |
| GLM-5.3 and GLM-5.3-Flash | The repository lists the 744B-A40B and 320B-A18B variants and paths for local serving | Check the version, precision and licence for the artefact you download |
The official Qwen project has also published a report on Qwen3.8-Omni, covering work with text, images, audio and video. Publishing a report and tools does not by itself confirm that all the models described can be downloaded and run locally. For pricing we need real weights and a working engine.
Smaller models also go into the comparison
It is worth keeping Gemma 4 12B, Mistral Small 4 and gpt-oss on the list. They represent different sizes and architectures. The name Small does not describe the needs of a home computer: Mistral Small 4 has 119B parameters in total and 6.5B active.
For Polish documents we add Bielik and PLLuM. The comparison with Qwen describes specific variants and licence limits; Bielik's terms of use (version 1.1 of 31 July 2025) forbid processing any personal data. Polish in a model's name or materials is a reason to include it in the trial. We check the quality on company terminology with real questions.
Open source, open weights and private data
Access to the weights lets you choose where the model runs. The licence sets what use, modification and distribution are allowed. The server code may have a different licence from the model itself. In the project we record the model ID, the revision and the terms that apply to the file in use.
Privacy also depends on the app. A local model connected to a cloud OCR service still sends the document to an outside service. Embeddings, backups and diagnostics need a separate check. The private LLM deployment page covers these parts together.
How we choose an open-weight model at Syntalith
For a help desk we measure whether cases are assigned correctly and how many corrections are needed. For a law firm, correct references to clauses and a refusal when a document is missing matter. In a code repository we check a working change and the tests. One general ranking cannot answer all these questions.
We start with a small set of candidates that fit the budget. They get the same material and criteria. The comparison covers the result, the latency at the target load and the cost of a correctly finished case. A model that needs fewer corrections can justify a higher compute cost.
If you want to choose a model for your company or your own computer, send us the use, the data requirements and the hardware specification. Syntalith will prepare the scope of the comparison, and can then handle installation, hosting and tuning of the chosen variant.
What an app on an open-weight model costs and when it pays back
- Today
- 400 documents × 12 min = 80 h × €34.90 = €2,792
- With the app, an employee checks the output in 25% of the time
- €698
- Hetzner GEX45 GPU server
- €214
- Saving
- €1,880 a month
Payback of €3,000
1.6 months
The volume, the time and the checking share are the scenario's assumptions, which we measure in the trial. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat. Price of the GEX45 with 24 GB of GPU memory from Hetzner's website on 22 October 2026, without the one-off setup fee of €209.
- Where the app takes documents from and where it saves results
- What stays in logs, backups and the history of questions
- Who has access to questions, answers and summaries
- Who updates the model, the app and the set of documents
Where to start
In the free process review we choose one task and 20 to 100 cases the team has already handled, and within 2 working days you get a written summary with a price. We roll out the local model as part of private LLMs, and the app around it as an AI app.
Worked example · Scenario
How we count the payback: an open-weight model
We run the same count for your team in the free process review.
€4,188
a month for the work on 600 questions
EU cloud version payback: 0.8 mo.
Our assumptions are 12 minutes to find an answer in the documentation, and checking in 15% of the old time. €34.90 an hour is the EU average hourly labour cost for 2025 according to Eurostat, for companies with 10 or more employees. Model cost: Claude Sonnet 5.5, about 6,000 + 500 tokens per answer in the EU cloud version, at the ECB rate of 1.1205 USD per euro on 22 October 2026. We count the payback for an app with a model in the EU cloud (from €3,000 net). We price the model on your server after a call, and the cost of people's work stays the same.
See the full calculation →
