# DeepSeek vs Llama vs Qwen

> Source: https://aiwiki.ai/wiki/deepseek_vs_llama_vs_qwen
> Updated: 2026-08-02
> Fact-checked: 2026-08-02
> Categories: AI Models, Large Language Models, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "DeepSeek vs Llama vs Qwen." aiwiki.ai, 2 Aug 2026. https://aiwiki.ai/wiki/deepseek_vs_llama_vs_qwen
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

[DeepSeek](https://aiwiki.ai/wiki/deepseek), [Llama](https://aiwiki.ai/wiki/llama), and [Qwen](https://aiwiki.ai/wiki/qwen) are families of large language models, not single systems. DeepSeek announced V4 as a preview on April 24, 2026 [1]. This comparison uses documentation dated or versioned no later than July 28, 2026 and selects six downloadable checkpoints: DeepSeek-V4-Pro and DeepSeek-V4-Flash [2][3][4], Llama 4 Scout and Maverick [5][6], and Qwen3.6-35B-A3B and Qwen3.6-27B [7][8][9]. The selection defines this article's scope; it is not a census of every release available by that date.

Downloadable checkpoints and hosted services are separate records. DeepSeek's dated announcement names both V4 API endpoints and downloadable weights [1]. Alibaba Cloud's pre-cutoff release log records hosted Qwen3.7 Max and Plus endpoint snapshots, while the downloadable Qwen records selected here are Qwen3.6 [7][10]. That separation does not establish the weight availability or license of Qwen3.7.

The six selected checkpoints can be compared directly on documented architecture, parameter counts, interfaces, licenses, and deployment examples. Their vendor benchmark tables do not form a matched leaderboard because the documented prompts, tools, reasoning budgets, context settings, precision, graders, and task versions differ [2][3][5][8][11].

## Selected checkpoints

| Checkpoint | Architecture | Parameters, total / active | Documented context | Documented input | Weight license |
| --- | --- | ---: | --- | --- | --- |
| DeepSeek-V4-Pro | [Mixture of Experts](https://aiwiki.ai/wiki/mixture_of_experts) language model | 1.6T / 49B | 1M tokens | Native image or video input is not specified in the cited checkpoint documentation | [MIT License](https://aiwiki.ai/wiki/mit_license) [2] |
| DeepSeek-V4-Flash | Mixture of Experts language model | 284B / 13B | 1M tokens | Native image or video input is not specified in the cited checkpoint documentation | MIT License [3] |
| Llama 4 Scout | Mixture of Experts with 16 experts and early-fusion multimodality | 109B / 17B | Supports 10M tokens; Meta says pre-training and post-training used 256K | Text and images | Llama 4 Community License [5][6] |
| Llama 4 Maverick | Mixture of Experts with 128 experts and early-fusion multimodality | 400B / 17B | 1M tokens | Text and images | Llama 4 Community License [5][6] |
| Qwen3.6-35B-A3B | Mixture of Experts with a vision encoder | 35B / 3B | 262,144 native; optional YaRN extension to 1,010,000 total input plus output | Text, images, and video examples | [Apache License 2.0](https://aiwiki.ai/wiki/apache_license) [8] |
| Qwen3.6-27B | Dense language model with a vision encoder | 27B / no separate active count reported | 262,144 native; optional YaRN extension to 1,010,000 total input plus output | Text, images, and video examples | Apache License 2.0 [7][9] |

Total and active parameters are different quantities. An expert-routed model stores its total parameters, while each token uses only a subset. The active count therefore does not equal checkpoint size or memory use, and it should not be compared as if it were the total size of a dense model [6].

## Architecture and training disclosures

DeepSeek describes both V4 checkpoints as expert-routed language models. Its technical report says the series combines Compressed Sparse Attention and Heavily Compressed Attention, uses manifold-constrained hyper-connections, and applies the Muon optimizer. The report states that Flash was pretrained on 32 trillion tokens and Pro on 33 trillion. These are developer disclosures, not independent measurements of efficiency or task quality [4].

Meta describes Llama 4 as an autoregressive expert-routed architecture with early fusion for text and image representations. Scout has 16 experts; Maverick has 128. Both activate about 17 billion parameters for a token, but Scout has 109 billion total parameters and Maverick has 400 billion [5][6].

Qwen3.6-35B-A3B combines a vision encoder with an expert-routed language model and reports 35 billion total parameters with 3 billion active. Qwen's repository identifies the 27B checkpoint as a dense model, while its card reports 27 billion parameters without a separate active count. Both cards document a vision encoder [7][8][9].

These disclosures do not by themselves establish relative output quality. Architecture, parameter count, and training-token count are inputs to an evaluation, not substitutes for one [11].

## Context and modalities

A maximum [context window](https://aiwiki.ai/wiki/context_window) is not the same as trained sequence length, retrieval accuracy, usable output budget, latency, or cost. DeepSeek documents a one-million-token window for both selected V4 checkpoints. Meta lists one million tokens for Maverick and support for 10 million tokens for Scout, but its launch article says Scout was pre-trained and post-trained at 256K and uses length generalization to reach the larger supported window [2][3][5][6].

The Qwen3.6 cards distinguish a native 262,144-token window from an optional YaRN configuration extending the total input plus output length to 1,010,000. They also warn that static YaRN can affect performance on shorter text and recommend enabling it only when long contexts are needed [8][9]. Nominal maxima should therefore not be ranked without the same long-context tasks, output allowance, and serving configuration.

The cited Llama 4 card lists text and images as inputs and text or code as outputs. The Qwen3.6 cards include text, image, and video examples. The reviewed DeepSeek V4 checkpoint materials describe language models but do not provide a comparable native image or video interface specification [2][3][5][8][9]. That documentation gap does not prove that every hosted wrapper or derivative built around V4 is unable to process other media.

## Reasoning controls

DeepSeek's model cards document three inference modes for the selected checkpoints: Non-think, Think High, and Think Max [2][3]. The Qwen3.6 cards document thinking and non-thinking operation, including controls for disabling the default thinking behavior [8][9].

The cited Llama 4 card does not document a directly equivalent reasoning-effort selector [5]. This is an interface difference in the reviewed records, not evidence that Llama cannot perform multi-step tasks or that no third-party provider offers additional controls.

## Weight access and licenses

All six checkpoints have downloadable [weights](https://aiwiki.ai/wiki/open_weights), so this article describes them as open-weight. That term does not make their legal terms or disclosure levels equivalent. The separate DeepSeek V4 Pro and Flash repositories and weights use the MIT License [2][3]. The two Qwen3.6 repositories use the Apache License 2.0 [7][8][9]. Llama 4 uses the custom Llama 4 Community License [5][12].

The Llama license includes conditions for redistribution, attribution, acceptable use, and some models created or improved using Llama materials or outputs. Its additional commercial term is narrower than a generic user threshold: if, on the Llama 4 version release date, products or services made available by or for the licensee or its affiliates had more than 700 million monthly active users in the preceding calendar month, the licensee must request a license from Meta before exercising the agreement's rights [12]. This is a factual summary, not legal advice.

Weight availability alone also does not establish compliance with the Open Source Initiative's system-level definition. That definition calls for the ability to use, study, modify, and share the system, together with data information, code, and parameters in the preferred form for modification [13]. Exact artifact licenses and disclosures should be checked for the checkpoint being deployed.

## Deployment implications

Model cards do not provide one universal hardware minimum. Requirements vary with weight precision, runtime, tensor parallelism, batch size, context allocation, and whether the system is self-hosted or served through an API [5][8][14].

Meta provides concrete deployment examples for Scout. Its official runtime README says Scout inference uses two GPUs with 80GB each in FP8, or one 80GB GPU in Int4. The model card separately says Scout can fit a single H100 GPU with on-the-fly Int4 quantization, while Maverick's FP8 weights fit a single H100 DGX host [5][14]. A DGX host is not the same claim as one H100 GPU.

The cited DeepSeek and Qwen cards give checkpoint precision and deployment instructions but do not establish comparable universal minima. Total parameter count, quantization, and the desired context should be included in any memory estimate; active parameters alone are insufficient [2][3][8][9].

## Comparing performance, cost, and safety

The developers publish benchmark tables, but those tables use different experimental conditions. DeepSeek's cards vary shot counts and reasoning modes [2][3]. Meta says all reported Llama 4 evaluations and testing in its model card used BF16 models [5]. Qwen documents an internal agent scaffold, context limits, sampling settings, task-set corrections, and external graders for different evaluations [8]. Copying selected values into one score table would hide those differences.

HELM's evaluation framework explains the general requirement: comparisons should hold scenarios and adaptation procedures constant where possible, report multiple metrics, and expose limitations. Prompt and adaptation choices can materially change results for the same model and task [11]. A useful local evaluation should fix the exact checkpoint or dated endpoint, prompt, tools, reasoning budget, context, precision, sampling, retries, task set, grading rule, and number of runs.

Price also requires a workload rather than a family name. Alibaba Cloud's July 15, 2026 pricing documentation shows that rates within one provider vary by exact model, service region, input-token tier, input versus output, cache status, batch mode, and promotion [15]. It does not provide a matched three-family cost study. Self-hosted cost also depends on hardware, utilization, engineering, and energy. A dated total-cost comparison should use the same workload and service assumptions for every option.

The reviewed Llama card and HELM framework do not supply a matched safety evaluation of all six checkpoints [5][11]. The amount of safety documentation is not a safety score. Deployment decisions require testing the exact checkpoint and surrounding system for the application's reliability, privacy, prompt-injection, misuse, and failure risks.

## Choosing among the checkpoints

The documented differences support conditional choices rather than one family-wide ranking:

- For image input, the selected Llama and Qwen cards document native interfaces. The Qwen cards also show video examples. The cited DeepSeek cards do not document a comparable native image or video interface [2][3][5][8][9].
- For long prompts, distinguish native length, trained length, optional extension, input and output allocation, and measured task accuracy. Scout's 10M support and Qwen's YaRN extension have different qualifications [5][6][8][9].
- For local deployment, compare the full checkpoint footprint and precision with available hardware. Active-parameter count alone is not a memory estimate [5][6][14].
- For redistribution or commercial use, review the exact MIT License, Apache License 2.0, or Llama 4 Community License terms that apply to the selected artifact [2][3][7][8][9][12].
- For task quality, latency, cost, and safety, run the same evaluation protocol on the exact checkpoints or dated hosted endpoints under consideration [11].

## References

[1] [DeepSeek V4 Preview Release](https://api-docs.deepseek.com/news/news260424), DeepSeek API Docs, April 24, 2026.

[2] [DeepSeek-V4-Pro model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/b5968e9190ef611bbf34a7229255be88a0e937c1/README.md), DeepSeek-AI, pinned revision.

[3] [DeepSeek-V4-Flash model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/blob/60d8d70770c6776ff598c94bb586a859a38244f1/README.md), DeepSeek-AI, pinned revision.

[4] [DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence](https://arxiv.org/abs/2606.19348v1), DeepSeek-AI et al., arXiv:2606.19348v1.

[5] [Llama 4 model card](https://github.com/meta-llama/llama-models/blob/038acb9fe1e1ddc5cf1b3989fb07b311a0ffcae4/models/llama4/MODEL_CARD.md), Meta, pinned revision, April 5, 2025.

[6] [The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation](https://ai.meta.com/blog/llama-4-multimodal-intelligence/), Meta, April 5, 2025.

[7] [Qwen3.6 repository README](https://github.com/QwenLM/Qwen3.6/blob/f1443092c29978643fd041ebe959676259e934f1/README.md), Qwen Team, pinned revision, April 22, 2026.

[8] [Qwen3.6-35B-A3B model card](https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/995ad96eacd98c81ed38be0c5b274b04031597b0/README.md), Qwen Team, pinned revision, April 24, 2026.

[9] [Qwen3.6-27B model card](https://huggingface.co/Qwen/Qwen3.6-27B/blob/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/README.md), Qwen Team, pinned revision, April 24, 2026.

[10] [Model lifecycle and updates](https://www.alibabacloud.com/help/en/model-studio/newly-released-models), Alibaba Cloud Model Studio, updated July 14, 2026.

[11] [Holistic Evaluation of Language Models](https://arxiv.org/abs/2211.09110), Percy Liang et al., Transactions on Machine Learning Research, 2023.

[12] [Llama 4 Community License Agreement](https://github.com/meta-llama/llama-models/blob/5fdf83110cc9daa7435dfba6eb304892cc0041b8/models/llama4/LICENSE), Meta, effective April 5, 2025.

[13] [The Open Source AI Definition, version 1.0](https://opensource.org/ai/open-source-ai-definition), Open Source Initiative, released October 28, 2024.

[14] [Meta llama-models README](https://github.com/meta-llama/llama-models/blob/0e0b8c519242d5833d8c11bffc1232b77ad7f301/README.md), Meta, pinned revision, October 10, 2025.

[15] [Model pricing](https://www.alibabacloud.com/help/en/model-studio/model-pricing), Alibaba Cloud Model Studio, updated July 15, 2026.