Comparison

Qwen2.5 vs DeepSeek-V3 vs Mistral Small 3.1

A side-by-side look at one named release from each of three open-weight families: Qwen2.5 (with Qwen2.5-7B-Instruct as the example release), DeepSeek-V3 (Base and Chat) and Mistral Small 3.1 24B Instruct (2503). Every row applies only to the release it names — these are not the families' current flagships — and the rows are based on the published model cards and licence pages associated with them.

At a glance

Three open-weight families, compared at one named release each: Qwen2.5 under Apache-2.0 (all open models except the 3B and 72B variants), DeepSeek-V3 with MIT for its repository code and the DeepSeek Model License for its Base and Chat weights, and Mistral Small 3.1 24B Instruct under Apache-2.0. Licence terms are set per release, not per family, so each row states the model it was read from. Sourced rows carry a per-product verification state.

Quick answer

Qwen, DeepSeek, Mistral can be used free of charge to start.

Where they differ — the rows worth reading first: Licence.
Identical on the record: Business model, Where it runs, Works offline, Runs on, Launched.

Qwen2.5 vs DeepSeek-V3 vs Mistral Small 3.1
Dimension	Qwen	DeepSeek	Mistral
Company	Alibaba	DeepSeek	Mistral AI
Category	Open Models	Open Models	Open Models
Business model	Open (open weights/source)	Open (open weights/source)	Open (open weights/source)
Licence	Apache-2.0 (Qwen2.5 open models except the 3B and 72B variants); later Qwen lines not read	MIT (DeepSeek-V3 repository code); DeepSeek Model License (DeepSeek-V3 Base and Chat weights)	Apache-2.0 for Mistral Small 3.1 24B Instruct; other releases differ (Mistral Medium 3.5: Modified MIT; Mistral Large 4: 'Open'; Voxtral TTS: CC BY-NC 4.0)
Pricing	Open-weight Qwen models are free to download; hosted access through Alibaba Cloud is usage-priced.	The chat assistant is free; the developer API is usage-priced, and open-weight models can be self-hosted without license fees. Pricing ↗	Open-weight models are free to download; hosted API access and commercial models are usage-priced. Pricing ↗
Documentation	—	—	Documentation ↗
Launched	2023-08	2023-11	2023-09
Availability	Global	Global	Global
Model compared (the release this page reads)	Qwen2.5 — Qwen2.5-7B-Instruct is the example release; the line also ships 0.5B–72B dense models plus Qwen2.5-Coder and Qwen2.5-Math Source ↗ Verified 2026-10-13	DeepSeek-V3 — read in its Base and Chat variants Source ↗ Verified 2026-10-13	Mistral Small 3.1 24B Instruct — Hugging Face release name Mistral-Small-3.1-24B-Instruct-2503 Source ↗ Verified 2026-10-13
Licence and commercial use (model compared)	Apache-2.0 for the Qwen2.5 open models except the 3B and 72B variants; commercial use allowed Source ↗ Verified 2026-10-13	DeepSeek-V3 only: MIT for the repository code, DeepSeek Model License for the Base and Chat weights; commercial use is supported subject to that licence's use-based restrictions Source ↗ Verified 2026-10-13	Mistral Small 3.1 24B Instruct only: Apache-2.0 open weights for commercial and non-commercial use. Other Mistral releases differ (Mistral Medium 3.5: Modified MIT; Mistral Large 4: 'Open'; Voxtral TTS: CC BY-NC 4.0) Source ↗ Verified 2026-10-13
Scale and architecture (model compared)	Dense Qwen2.5 family from 0.5B to 72B parameters Source ↗ Verified 2026-10-13	671B total parameters with 37B activated per token (mixture-of-experts) Source ↗ Verified 2026-10-13	24B dense model (Mistral Small 3.1 24B Instruct) Source ↗ Verified 2026-10-13
Context window (model compared)	131,072 tokens for Qwen2.5-7B-Instruct, with 8,192-token generation Source ↗ Verified 2026-10-13	128K tokens (DeepSeek-V3) Source ↗ Verified 2026-10-13	128k tokens (Mistral Small 3.1 24B Instruct) Source ↗ Verified 2026-10-13
Modality (model compared)	Text (Qwen2.5 language models) Source ↗ Verified 2026-10-13	Text (DeepSeek-V3) Source ↗ Verified 2026-10-13	Text and image (Mistral Small 3.1 adds vision understanding) Source ↗ Verified 2026-10-13
Family breadth (context only — not the release compared)	General, Coder and Math lines from 0.5B to 72B; later Qwen lines are outside this page's licence reading Source ↗ Verified 2026-10-13	The DeepSeek-V3 release ships Base and Chat variants; no other DeepSeek release is covered here Source ↗ Verified 2026-10-13	Generalist, code, OCR, audio and moderation lines; licences differ by release, so the Apache-2.0 row above applies to Mistral Small 3.1 only Source ↗ Verified 2026-10-13
Updated	2026-10-13	2026-10-13	2026-10-13

Fields taken from the product records use the record dates shown. Sourced claims are labelled per product: verified against the linked page on the date shown, taken from the product record without re-checking, or not independently verified.

When to choose each

Qwen — When you want the Qwen2.5 line: dense models from 0.5B to 72B under Apache-2.0 (except the 3B and 72B variants), with long context and broad language coverage.
DeepSeek — When you want the DeepSeek-V3 release: a large mixture-of-experts model with published Base and Chat weights whose licence explicitly supports commercial use.
Mistral — When you want the Mistral Small 3.1 24B Instruct release: Apache-2.0 open weights at a mid size, with vision input. Other Mistral releases use different terms.

Each tool on the record

Qwen

Qwen is made by Alibaba. The record lists it as Open (open weights/source), first recorded in 2023, and fit for Running models yourself. Its values below carry their own source and date.

Qwen is listed under Open Models: Open-weight and open-source model families and their hosted services.

Company: Alibaba
Category: Open Models
Business model: Open (open weights/source)
Licence: Apache-2.0 (Qwen2.5 open models except 3B and 72B; later Qwen lines not read)
Launched: 2023-08
Availability: Global
Tasks it fits: Running models yourself
How you use it: Model weights
Runs on: Not recorded
Where it runs: Local machine
Works offline: Partly
Commercial use: Permitted

Pricing on the record: Open-weight Qwen models are free to download; hosted access through Alibaba Cloud is usage-priced.

Sources on the record: Project repository, Qwen (Wikipedia), Qwen2.5 release notes, Qwen2.5-7B-Instruct model card (Hugging Face).

DeepSeek

DeepSeek is made by DeepSeek. The record lists it as Open (open weights/source), first recorded in 2023, and fit for Running models yourself. Its values below carry their own source and date.

DeepSeek is listed under Open Models: Open-weight and open-source model families and their hosted services.

Company: DeepSeek
Category: Open Models
Business model: Open (open weights/source)
Licence: DeepSeek-V3: MIT (repository code); DeepSeek Model License (Base and Chat weights)
Launched: 2023-11
Availability: Global
Tasks it fits: Running models yourself
How you use it: Model weights
Runs on: Not recorded
Where it runs: Local machine
Works offline: Partly
Commercial use: Permitted

Pricing on the record: The chat assistant is free; the developer API is usage-priced, and open-weight models can be self-hosted without license fees.

Sources on the record: Official website, DeepSeek (Wikipedia), Official pricing, DeepSeek-V3 model card (Hugging Face), DeepSeek-V3 code licence (LICENSE-CODE, MIT), DeepSeek-V3 model licence (LICENSE-MODEL, DeepSeek License Agreement v1.0).

Mistral

Mistral is made by Mistral AI. The record lists it as Open (open weights/source), first recorded in 2023, and fit for Running models yourself. Its values below carry their own source and date.

Mistral is listed under Open Models: Open-weight and open-source model families and their hosted services.

Company: Mistral AI
Category: Open Models
Business model: Open (open weights/source)
Licence: Apache-2.0 (Mistral Small 3.1 24B Instruct); other releases vary
Launched: 2023-09
Availability: Global
Tasks it fits: Running models yourself
How you use it: Model weights
Runs on: Not recorded
Where it runs: Local machine
Works offline: Partly
Commercial use: Permitted

Pricing on the record: Open-weight models are free to download; hosted API access and commercial models are usage-priced.

Sources on the record: Official website, Mistral AI (Wikipedia), Documentation - Mistral AI, Official pricing, Mistral Small 3.1 24B Instruct model card (Hugging Face), Models overview (licences per release).

Frequently asked questions

Which of these has been available longest?

Qwen — first recorded in 2023.

Which of these can I start using for free?

Free to start: Qwen, DeepSeek, Mistral.

Which of these is open source?

None of these is open source; every tool in this comparison is proprietary.

Which of these works offline?

None of these is recorded as working offline.

Which of these has a public API?

A public API is recorded for Qwen, DeepSeek, Mistral.

Which of these can be used commercially?

Commercial use is recorded as allowed for Qwen, DeepSeek, Mistral.

Which platforms do these run on?

Recorded platforms — Qwen: Not recorded · DeepSeek: Not recorded · Mistral: Not recorded.

Can any of these be self-hosted or run locally?

Recorded as self-hosted or local: Qwen, DeepSeek, Mistral.

Can I download and run these models on my own hardware?

Yes — all three releases compared here publish weights you can download and run yourself. Hardware needs vary widely with model size, from a 24B dense model to a 671B mixture-of-experts model.

Do these licences allow commercial use?

It depends on the release, and code and weights can differ. Qwen2.5 (except the 3B and 72B variants) and Mistral Small 3.1 24B Instruct are Apache-2.0, which permits commercial use. DeepSeek-V3 is split: MIT for the repository code, and the DeepSeek Model License for the Base and Chat weights, which supports commercial use subject to its use-based restrictions. Other releases from these families were not read for this page and carry their own terms.

Are the licences shown here final?

No. Licences are set per model release, not per family, and can change between versions and variants. Each row here names the release it was read from; other models from the same families may use different terms.

How to use a comparison

Start from the task, not the brand: filter the index by what you need to get done.
Check the evidence: a value marked “verified” names the source page and the date it was checked.
Prefer the vendor’s own page for the final decision: prices and limits change without notice.

Related guides

Open-source AI models

Sources and record dates

Record dates describe index updates, not independent product tests. Check linked documentation for current features and pricing.