The Open-Weight Models You Can Actually Run Yourself: DeepSeek, Mistral, and the Self-Hosting Option

Somewhere between a hospital that can’t send patient data to an outside API and a hobbyist who just wants to see how a model thinks, there is a growing category of AI that never phones home at all. This is the second article in LIWARSE’s five-part series on the AI landscape by type of use.

Our provider series already covered the open-weight releases from OpenAI (gpt-oss), Meta (Llama 4), xAI (open Grok weights), and Alibaba (Qwen). This article covers the two other labs whose open models a member of the public, a small clinic, or an independent researcher is most likely to actually download and run today: DeepSeek and Mistral.

DeepSeek: Free, Fast, and Fully Downloadable

DeepSeek’s V4 and V4-Pro models, released in April 2026, followed the free-app strategy that made the company’s earlier R1 model the most downloaded app in the U.S. App Store within a week of its January 2025 launch. The underlying weights are released under the permissive MIT license, meaning anyone can download, modify, and redeploy them with no royalty owed back to the company — a genuinely unguarded release by LIWARSE’s definition. In July 2026 DeepSeek shipped V4-Flash, a retrained variant tuned specifically for coding and agentic tool use that, unusually, now outperforms the company’s own larger V4-Pro on those tasks.

DeepSeek’s chat app itself is proprietary and hosted in China, subject to Chinese data-handling law — a fact any clinician or researcher outside China should weigh before typing patient or proprietary data into it. The downloadable weights carry no such constraint once run on a user’s own hardware, which is precisely the point: the model and the app built on top of it are two different trust decisions.

---

Mistral: Europe’s Open-Weight Flagship

Mistral AI, based in Paris, has built its identity around shipping genuinely competitive open-weight models under the Apache 2.0 license — among the most permissive terms available, allowing unrestricted commercial reuse. Its flagship, Mistral Large 3 (December 2025), is a sparse mixture-of-experts model with 675 billion total parameters, of which only 41 billion activate for any given query — a design that keeps inference costs manageable while retaining frontier-scale capacity. A smaller Mistral Small 4 targets developers who need to self-host on more modest hardware.

Mistral’s consumer assistant, rebranded from Le Chat to Vibe in 2026, sits on top of these models the way ChatGPT sits on top of GPT — but the open weights underneath mean a European hospital system with strict data-residency rules can, in principle, run an equivalent model entirely inside its own infrastructure rather than trust a third-party app. That combination — competitive capability plus EU data residency plus a genuinely open license — has made Mistral a default recommendation for institutions bound by GDPR-grade compliance.

---

Risks and Benefits Through the LIWARSE Lens

Benefits

• Local deployment lets clinics, researchers, and public institutions keep sensitive data off third-party servers entirely — a direct, practical service to patient privacy and data sovereignty.
• Free, unrestricted licenses (MIT for DeepSeek, Apache 2.0 for Mistral) put frontier-adjacent capability within reach of institutions and individuals who could never afford flagship API pricing at scale.
• Mistral’s efficient mixture-of-experts design lowers the hardware bar for self-hosting, widening who can realistically run a capable model themselves.

Risks

• Once downloaded, both model families carry only whatever safety tuning was applied before release — a layer any sufficiently motivated user can strip away, exactly the hazard LIWARSE’s Open Source AI Without Guardrails position paper describes.
• DeepSeek’s free hosted app is a separate trust surface from its open weights; users who assume “open” means “private” may unknowingly route sensitive queries through Chinese servers instead of their own hardware.
• Neither lab has published independent, third-party tamper-resistance verification of the kind LIWARSE’s Negative Intelligence framework calls for as standard practice.

The LIWARSE Assessment

DeepSeek and Mistral occupy the same structural position in LIWARSE’s framework as the open releases already covered in the provider series: genuinely useful, genuinely unguarded once downloaded, and safe only to the extent that the person running them chooses to keep the safety tuning intact. For a solo researcher or an under-resourced clinic, that trade can be the right one. For anyone deploying at scale, LIWARSE’s standing recommendation applies here too — pair any open-weight deployment with an independent, policy-based screening layer the operator controls, rather than relying on whatever guardrails shipped in the download.

---

Under the 3 Absolute Laws, an open model is not unsafe by nature — it simply moves the burden of safety from the provider to whoever presses run. DeepSeek and Mistral have made that burden light enough for almost anyone to carry. Whether they carry it responsibly is now a question for each downloader, not each company.

— The LIWARSE Movement | liwarse.org
Safety of Life · Advancement of Life · Together.