DeepSeek vs Qwen (2026): which one should you use?

Corrected 2026-09-29: an earlier version listed Outlier's Vision tier as Qwen3.6-35B-A3B (19.0 GB), the model it replaced; Vision 3.8 is built on Qwen3.8-27B, a 15.5 GB download.

Most comparisons of these two open 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 labs publish downloadable checkpoints while also selling hosted access to models they never release.

Disclosure: I build Outlier, a Mac app that runs models on-device, and five of its six tiers are Qwen weights. The sixth is Gemma. I'm not a neutral party. Check the Qwen-favorable claims against the sources at the bottom, and I've been 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 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 weights are under the MIT License. Qwen's Qwen3.5 and Qwen3.6 models and Qwen3.8-27B are Apache 2.0 too: the Hugging Face model cards list it, and the Qwen repository points to the license file released with the weights. Both licenses are OSI-approved, with neither imposing a use policy or an approval form. One honest caveat on the Qwen side: that Apache 2.0 sweep holds for the models the app ships, not for every checkpoint Alibaba has released. Qwen3.8-2.4T-A95B ships under its own Qwen3.8-Max License, which lets you use, modify and sell it but asks you to display the model name in products above 100 million monthly active users or US$20 million in monthly revenue, and requires a separate license for model-as-a-service or AI work assistant businesses above US$50 million in revenue. Qwen3.8-Flash-Next, an open-weight model in the same generation, is under a third license: its model card names the Qwen Community License 1.0, which repeats the 100 million user and US$20 million monthly revenue thresholds for showing the model name and adds its own separate-license condition for model-as-a-service and AI work assistant businesses, so read it before you build one. Qwen2.5-72B is under Alibaba's custom Qwen License Agreement, which makes you request a license above 100 million monthly active users, and Qwen2.5-3B is under a Qwen Research License for non-commercial use, so check the license on the specific model you pull, not just the family.

Compare the alternatives. Meta's Llama Community License (the Llama 3.3 version asks you to request a license from Meta above 700 million monthly active users) and Google's Gemma Terms of Use are custom vendor licenses with acceptable-use policies attached, and neither appears on the Open Source Initiative's approved list. Those conditions travel downstream. I can speak firsthand to the Gemma side: our Quick tier is a Gemma 4 model, and Google publishes Gemma 4 under Apache 2.0 (see the model card); the Gemma Terms of Use send Gemma 4 readers to a separate Gemma 4 license, and the custom terms apply to the earlier models they list.

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. That ladder lets an app match a checkpoint to a customer's hardware. 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."

DeepSeek builds at another scale. V4-Flash is 284B total with 13B activated, 1M 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, but the full weight set still has to live somewhere. DeepSeek has kept shipping since: DeepSeek-V4.1-Flash is open under MIT too, with 552B backbone parameters, 8B activated per token in prefill and 16B in decode, a one-million-token context and image input, and DeepSeek-V4-Flash-Vision-Exp is an experimental multimodal V4 model (about 305B parameters in its Safetensors metadata, also MIT). None of those is laptop-sized. Qwen has narrowed the size gap from its side: its README says Qwen3.8 brings a Qwen-Max-class model to open release, the 2.4T-A95B (2.4T total, 95B activated, 262,144 tokens of native context), and Qwen3.8-Flash-Next is an open-weight model whose card lists 125B language-model parameters, 6B activated and 262,144 tokens of native context. qwen3.7-max and qwen3.7-plus appear on Alibaba's Model Studio pricing page as hosted models billed per token.

One correction to the "DeepSeek is too big for a laptop" story, since it cuts against the recommendation below: DeepSeek's earlier R1 distill series (1.5B, 7B, 8B, 14B, 32B and 70B, all MIT) runs on consumer Macs: Ollama's library page for deepseek-r1 lists a 7b tag at a 4.7 GB download. The caveat is what they are: fine-tunes of Qwen2.5 and Llama 3.x models (the 7B card names Qwen2.5-Math-7B as its base, the 70B names Llama-3.3-70B-Instruct) carrying DeepSeek's reasoning traces, not DeepSeek architectures, and a generation behind the V4 line. But if you want DeepSeek-lineage weights on a 16 GB machine, they exist and the license is as clean as V4's.

What runs on a Mac, and at what RAM

These are the exact checkpoints Outlier ships, quantized, on Apple Silicon. Qwen3.5-4B: 2.37 GB, 6 GB RAM, 71.7 tok/s on an M1 Ultra and about 32 tok/s on an M4 MacBook Air (single prompt, 2026-05-04, indicative). 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. Qwen3.8-27B: 15.5 GB, 24 GB RAM. Qwen3.5-397B-A17B: 209 GB, 64 GB RAM, 1.59 tok/s, streaming experts off SSD at ~14 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 1.6T at any precision is past what I'd ship.

Strengths by task, and language coverage

On coding, Qwen is where I have numbers. Our Core tier, Qwen3.6-27B, measures 0.951 on HumanEval, and on a blind slice of SWE-bench Verified the local 27B came in 46% (23 of 50, build 1.11.757, 2026-08-09): our runs, our quantizations. We also published a 54-prompt side-by-side with Claude Opus so you can compare the answers yourself; Claude is still faster and stronger on the hardest work.

On breadth, Qwen is ahead: open Qwen3-VL for vision, Qwen3-Omni for text/audio/image/video in with text and audio out, plus ASR, TTS and embedding models. DeepSeek's open weights now take images too, through V4.1-Flash and the experimental Vision-Exp model, but only at the sizes above. Its generosity otherwise runs a different direction: DeepGEMM, DeepEP, FlashMLA, 3FS and DeepSeek-OCR are all MIT, infrastructure that 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 and both treat Chinese and English as first-class. If a lower-resource language decides it, test both yourself.

DeepSeek and Qwen, line by line

Where each family genuinely wins

DeepSeek wins on license simplicity at the frontier. Releasing a 1.6T-parameter checkpoint under plain MIT is unusual: Qwen's larger 2.4T flagship ships under its own license, with the user-count and revenue conditions described above.

DeepSeek's hosted prices undercut Alibaba's mid and top tiers. deepseek-flash is $0.15 per million input tokens on a cache miss off-peak and $0.30 at peak, and $0.003 on a cache hit ($0.006 at peak), with output at $0.60 ($1.20 at peak): the cache-hit path is fifty times cheaper, which matters for workloads that repeat a long prefix. deepseek-v4-pro runs $0.66 / $0.022 / $1.98 on the same axes off-peak, and double that at peak. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, except Chinese public holidays; every other hour is off-peak. DeepSeek's update notes say the name deepseek-flash now reaches DeepSeek-V4.1-Flash-1017, with the same API call and pricing. Alibaba lists qwen3.7-plus at $0.4 in / $1.6 out for requests up to 256K input tokens (it steps to $1.2 / $4.8 from 256K to 1M) and qwen3.7-max at $2.5 / $7.5; its newer qwen3.8-max lists $2 / $6 and qwen3.8-flash $0.15 / $0.47. Alibaba's own Flash tier lists lower rates still, from $0.03 in / $0.13 out for qwen3.7-flash on requests up to 32K input tokens.

DeepSeek documents its data controls. Its terms (last update March 27, 2026) describe an in-product "Improve the model for everyone" toggle and assign output rights: "We assign any rights, title, and interests—if any—in the Outputs of the Services to you." Its privacy policy (last update Feb 10, 2026) describes a right to opt out of training. I couldn't read Qwen's own privacy policy, which renders only with JavaScript, so I'm not comparing it here.

Qwen wins on the ladder, the modalities and the cadence. Eight open sizes in one quarter is an upgrade path, not a dead end. Qwen Code, its Apache-2.0 terminal agent, lists support for "OpenAI, Anthropic, Gemini, and Qwen APIs. Any third-party provider or local model (Ollama / vLLM)," and its privacy page says "Qwen Code itself does not use your prompts, code, or responses for model training"; what your chosen provider does with them is that provider's policy. Qwen Studio, at chat.qwen.ai, is a free assistant from the same lab.

Both carry caveats. The $2.5 / $7.5 and $0.4 / $1.6 figures above are Alibaba's list rates: its pricing page says it lists standard prices only and sends you to the Model Studio console for current promotions, and it currently labels qwen3.7-plus with a limited-time 20% off. Check the console for the live rate before you budget. DeepSeek reserves the right to adjust prices, and its rates now differ between peak and off-peak hours.

Which family fits your constraint

Pick DeepSeek if you run server-side inference at volume and its price per token against Alibaba's mid and top tiers matters, if you want frontier-scale open weights and the hardware to serve them, or if a documented training opt-out is something your compliance review looks for.

Pick Qwen if you need a current-generation open 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 free hosted assistant from the same lab to try before you download anything.

Pick neither hosted service if the data can't leave your jurisdiction. DeepSeek stores personal data in the PRC; Alibaba Cloud Model Studio is priced by region (Singapore, Germany, the US, mainland China and others), so check which one you'd call. Running the weights yourself is the only version where that question doesn't arise, which is why Outlier exists, and why it's Qwen inside. What we don't do: image generation, mobile, voice mode, Windows or Linux, cloud sync; web search is Pro-only and opt-in.