DeepSeek vs Qwen: open model families compared (2026)

Most comparisons of these two start with benchmark charts. The questions that actually decide things are duller: what license are the weights under, which sizes exist, and will any of them fit on the machine in front of you? Both are Chinese labs that publish downloadable checkpoints while also selling hosted access to models they never release — and they answer those three questions very differently.

Disclosure, because it matters here: I build Outlier, a Mac app that runs models on-device, and six of its seven tiers are Qwen weights. The seventh is Gemma. I'm not a neutral party. Check the Qwen-favorable parts against the sources at the bottom — and I've tried to be equally specific about where DeepSeek is the better answer, because on several axes it plainly is.

Licensing is the under-covered story

Licensing decides whether you can build a business on top of the weights, and hardly anyone writes about it. Both DeepSeek-V4-Pro and DeepSeek-V4-Flash state on their model cards that the repository and the weights are under the MIT License. Qwen is equally blunt: "All our open-weight models are licensed under Apache 2.0." Both are OSI-approved, and neither imposes a use policy or an approval form. Apache 2.0 asks for attribution, a copy of the license, and adds an explicit patent grant; MIT asks for even less.

Compare the alternatives you meet elsewhere in open weights. Meta's Llama Community License and Google's Gemma Terms of Use are custom vendor licenses with acceptable-use policies attached, and neither is open source in the OSI sense — those conditions travel downstream to whoever you redistribute to. I can speak firsthand to the Gemma side: our Quick tier is a Gemma model, and it's the only one of seven where shipping meant reading a use policy and passing terms along instead of dropping in a license header. For a downstream product, MIT and Apache 2.0 are the difference between "publish it" and "call a lawyer first."

Sizes: a ladder versus two very large rungs

Qwen's open line covers 0.8B, 2B, 4B, 9B, 27B, 35B-A3B, 122B-A10B and 397B-A17B, released in waves between February 16 and April 22, 2026. The ladder isn't cosmetic — it lets an app match a checkpoint to a customer's hardware instead of accepting one take-it-or-leave-it artifact. Qwen3.6-27B is the interesting rung: dense, 64 layers, a hybrid Gated DeltaNet plus Gated Attention architecture, 262,144 tokens of native context extensible to roughly 1,010,000 via YaRN. Its model card bills it as "Flagship-Level Coding in a 27B Dense Model," and 27B dense quantizes onto a 32 GB Mac.

DeepSeek builds at another scale entirely. V4-Flash is 284B total with 13B activated, 1M token context, mixed FP4/FP8 with the MoE experts in FP4. V4-Pro is 1.6T total with 49B activated and the same 1M context. Sparse MoE keeps activated parameters low, which is genuinely kind to bandwidth-bound machines, but the whole weight set still has to live somewhere. The asymmetry runs the other way too: Qwen's strongest models aren't open at all. qwen3.7-max, qwen3.7-plus and qwen3.8-max-preview are API-only, so the open family visibly trails its own flagship.

What runs on a Mac, and at what RAM

Here I can give measured numbers rather than estimates, because these are the exact checkpoints Outlier ships, quantized, on Apple Silicon. Qwen3.5-4B: 2.37 GB download, 6 GB RAM, 71.7 tok/s on an M1 Ultra and about 32 tok/s on an M4 MacBook Air. Qwen3.5-9B: 5.04 GB, 12 GB RAM, 53.4 tok/s. Qwen3.6-27B: 15.13 GB, 24 GB RAM, 20.7 tok/s. Qwen3.6-35B-A3B: 19.0 GB, 24 GB RAM. Qwen3.5-397B-A17B: a 209 GB download needing a 64 GB machine, running at 2.1 tok/s by streaming experts from SSD with paged MoE at roughly 11 GB peak RSS.

That last tier cuts against my own argument, so I'll say it plainly: expert streaming does let a 397B MoE run on a 64 GB desktop, so "too big for consumer hardware" isn't absolute. But 284B in mixed FP4/FP8 and 1.6T at any precision are past what I'd ship to someone expecting the app to feel usable. On a Mac today, DeepSeek's open weights are a rental-GPU proposition and Qwen's are a local one.

Strengths by task, and language coverage

On coding, Qwen is the family I can put numbers behind. Our Code tier — Qwen3.6-27B — measures 0.866 on HumanEval, and on a blind slice of SWE-bench Verified the local 27B came in around 45% (18 of 40). In a 54-prompt head-to-head against Claude Opus it matched the rubric on 98.9% of checks. Those are our runs on our quantizations, worth exactly as much as your trust in a founder publishing his own benchmark.

On breadth, Qwen is clearly ahead: open Qwen3-VL for vision, Qwen3-Omni for text/audio/image/video in with text and audio out, plus ASR, TTS, embedding and image generation models. DeepSeek's public generosity runs a different direction — DeepGEMM, DeepEP, FlashMLA, 3FS and DeepSeek-OCR are all MIT, and that infrastructure has quietly made other people's inference faster, local inference included.

On language coverage I'll be honest rather than useful. I have no primary source publishing a verified language list or an apples-to-apples multilingual eval for either family, and I won't invent one. Both labs are Chinese; both treat Chinese and English as first-class. If a specific lower-resource language decides your choice, test both yourself.

Side by side

Where each family genuinely wins

DeepSeek wins on license simplicity at the frontier. Releasing a 1.6T-parameter checkpoint under MIT is unusual. Nobody else at that scale hands over weights with a license that short.

DeepSeek wins on hosted price, decisively. deepseek-v4-flash is $0.14 per million input tokens on a cache miss and $0.0028 on a cache hit, with $0.28 output — the cache-hit path is fifty times cheaper than the miss, which matters a great deal for workloads that repeat a long prefix. deepseek-v4-pro runs $0.435 / $0.003625 / $0.87 on the same axes.

DeepSeek wins on user data control. Its terms document an in-product "Improve the model for everyone" toggle plus an email opt-out, and explicitly assign output rights: "We assign any rights, title, and interests—if any—in the Outputs of the Services to you." Qwen documents no equivalent toggle; its policy lists training on de-identified User Content and Feedback under a legitimate-interests basis, offering only generic rights to object or withdraw consent through a DPO email.

Qwen wins on the ladder, the modalities and the cadence. Eight open sizes across several waves in one quarter means an upgrade path rather than a dead end. Qwen Code, its Apache-2.0 terminal agent, deliberately speaks to "OpenAI, Anthropic, Gemini, and Qwen APIs. Any third-party provider or local model (Ollama / vLLM)," and states it doesn't train on your prompts or code. Qwen Studio ships native clients on Web, iOS, Android, macOS and Windows.

Both carry caveats worth pricing in. Alibaba's headline rates are limited-time discounts — 50% off qwen3.7-max, 20% off qwen3.7-plus — with no published end date, so budget for roughly double; the Qwen Code free tier ended 2026-04-15, and the new-user free quota is 90 days, Singapore region, real-time inference only. DeepSeek's docs mark the widely-integrated deepseek-chat and deepseek-reasoner names deprecated on 2026/07/24 15:59 UTC and reserve the right to adjust prices. As of mid-2026 I could not confirm an official DeepSeek macOS client from a primary source — its App Store listing covers iPhone and iPad only.

Who should pick which

Pick DeepSeek if you run server-side inference at volume and price per token dominates, if you want frontier-scale open weights and hardware to serve them, or if a documented training opt-out is something your compliance review looks for.

Pick Qwen if you need a model that fits consumer hardware, if Apache 2.0's patent grant matters to your legal team, if you want open vision or speech models rather than text alone, or if you want a desktop client today.

Pick either to fine-tune a commercial product. MIT is marginally less paperwork than Apache 2.0, and both are far less than a vendor community license.

Pick neither hosted service if the data can't leave your jurisdiction. DeepSeek stores personal data in the PRC; Qwen stores and processes it in Singapore and Mainland China. Running weights on your own machine is the only version where that question doesn't arise — which is why Outlier exists, and why it's Qwen inside. Equally plainly, what we don't do: no web search, no image generation, no mobile app, no voice mode, no Windows or Linux build, no cloud sync.