Qwen vs DeepSeek vs Llama: Which Should You Build On?

The metro out to Alibaba's Xixi campus, where the Qwen team works, takes about twenty minutes from my apartment, and DeepSeek's people sit a few districts the other way. So I get to watch this comparison move in something close to real time, and it moves fast. The current open checkpoints are Qwen3.8 and Qwen3.6, DeepSeek V4 Pro and Flash, and Llama 4 Scout or Maverick. Alibaba's hosted Qwen3.8-Max belongs in a different column, because renting it is a different product from downloading the weights it sits on. If you blur those versions together, the rest of the comparison is wrong before it starts.

Qwen vs Llama vs DeepSeek: which should you build on?

Pick by job and by deployment. Qwen3.6 is the breadth and ecosystem choice, with 27B dense and 35B-A3B open models under Apache-2.0. DeepSeek V4 Flash is the API-value choice; V4 Pro buys more capability at a still-low token price but is enormous to self-host. Llama 4 remains the Western-origin ecosystem choice under Meta's community licence. Stanford's 2.7% country-level gap from March 2026 is now historical context, not a live score, because every Chinese family in this paragraph has moved since then.

Key Takeaways- Compare exact checkpoints: Qwen3.6, DeepSeek V4 and Llama 4, not family names inherited from last year's releases.- Qwen3.6 = breadth and Apache-2.0; DeepSeek V4 = reasoning and API value; Llama 4 = Western-origin tooling and Meta's community licence.- Qwen is the most-downloaded open family, past 1 billion downloads and 200,000 derivatives by January 2026 by Alibaba's own count.- Qwen3.8 is Alibaba's newest open checkpoint, out since 8 August 2026 under a custom licence; Qwen3.8-Max is the hosted version built on it.

Everything below assumes you've already decided Chinese weights are on the table. If that's still the open question, I argue it out in my piece on whether China is winning the AI race, and the wider field sits in my guide to the Chinese open-source AI ecosystem.

How I'd rank them

Not by leaderboard. I rank by deployment reality, which means five criteria do the work, in roughly this order of weight.

• Capability: is it genuinely good enough for the job, under a comparable evaluation rather than a vendor-selected table?
• Cost: cost per token on an API, and GPU footprint when you self-host. This is where the real gaps open up.
• Licence: can you ship it commercially without a legal headache? Apache-2.0 and MIT-style terms win here.
• Ecosystem: how many fine-tunes, quantisations and ready variants exist? More means less work for you.
• Self-host friendliness: does it run cleanly on sane hardware with the open inference stack you already use?

Running these on my own GPUs cured me of leaderboard worship faster than any argument could. The model that wins a project is the one that fine-tunes cleanly on your data and costs least to serve, not the one with a marginally higher score this week. Two of my last three client builds shipped on a model that wasn't top of any chart.

One structural note before the table. Hosted previews and downloadable weights are different products, and August 2026 makes the point neatly: Qwen3.8-Max launched as a hosted model on 3 August and the open Qwen3.8 weights followed on 8 August, under a custom licence rather than Apache-2.0. So a ranking that mixes "what you can download" with "what you can rent" will mislead you. The table below is the downloadable side; the hosted flagships get their own column in the prose.

The three families at a glance

All three are genuinely production-grade. Where they differ is licence and cost posture, far more than capability. Qwen passed 1 billion cumulative downloads and 200,000 derivative models on Hugging Face in January 2026, by Alibaba's own count, which makes it the largest open family by adoption. Here's how I'd summarise the three at a working level.

Adoption more broadly has shifted the same way. Chinese open-weight models have overtaken US ones in total Hugging Face downloads, and Qwen-derived variants alone were more than 40% of new derivatives by August 2025. That matters for builders in a practical way: more derivatives means a quantisation or domain fine-tune that fits your latency budget probably already exists, so you skip a training run entirely.

One caution on that table. Licences move between versions, and a family-level summary is a starting point rather than gospel. Open the model card and read the actual licence for the exact checkpoint you plan to ship.

Capability: how close are they really?

Closer than the marketing wars suggest, but no longer reducible to one number. The 2026 Stanford AI Index recorded a 2.7% gap between the top US and top Chinese models on Chatbot Arena in March, on the Index's technical performance chapter. Since then DeepSeek V4, Qwen3.6 and the hosted Qwen3.8-Max have arrived, so that figure cannot honestly be written in the present tense. For most business workloads, capability is a gate rather than the final selector. Cost, serving constraints and licence usually break the tie.

[CHART: grouped-bar comparing Qwen, DeepSeek and Llama across four qualitative axes]

The working distinction is clearer when it is versioned. DeepSeek V4 leans towards reasoning and code. Qwen3.6 is the open all-rounder and easiest of the three to size for a specific deployment. Llama 4 is the Western-origin option with the surrounding Meta ecosystem, though it's worth knowing that Meta's frontier effort has moved to its closed Muse line, with Llama 4 in maintenance mode and the larger Behemoth paused. The tooling around Llama is still deep, but I wouldn't plan a multi-year bet on the next open Llama checkpoint.

Benchmark obsession is actively misleading for builders when the rows come from different provider harnesses. The model that wins your project is the one that clears your quality threshold, behaves on your data and fits your serving budget, not the one with a marginally higher self-reported score.

So I treat capability as a near-tie and go straight to the things that genuinely differ.

Author's directional assessment on a 1 to 5 scale, not benchmark numbers. Scores reflect the August 2026 releases: DeepSeek V4, Qwen3.8 and 3.6, and Llama 4 with Meta's open line in maintenance mode.

Cost and licensing: the real differentiators

This is where the decision actually turns. On DeepSeek's official API, V4 Flash costs $0.14 per million cache-miss input tokens and $0.28 per million output tokens; V4 Pro costs $0.435 and $0.87. Qwen3.8-Max is sold through Alibaba's hosted services at around $2 and $6 respectively, normal hosted-token economics rather than open-weight ones. Llama has no first-party Meta token API in this comparison, so its cost is your host's price or your own infrastructure.

The families split cleanly here. Qwen3.6 is Apache-2.0. DeepSeek V4 is MIT. Llama 4 uses Meta's community licence, with use restrictions that need an actual legal read. The important Qwen footnote is that the open line moved off Apache-2.0 at 3.8: the new weights carry a custom Qwen licence with revenue-based commercial clauses, so 3.6 remains the Apache-2.0 answer.

When I self-host, the Apache-2.0 question is the first thing I check, because a clean answer removes an entire category of legal review. With Llama I've had to stop and read the community licence properly before a client deployment, and twice that clause changed the answer. That friction is real, and it's why Qwen tends to win my "just ship it" projects.

Self-hosting compounds the cost story. A smaller-footprint model means less GPU, which means a smaller bill every month rather than a one-off saving at launch. I work that maths through in my guide to running open models on-prem versus cloud, and the governance side of the same decision is in my piece on whether DeepSeek is safe for enterprise.

On the adoption side, the cost posture is doing what you'd expect. Chinese models took the majority of OpenRouter's top-10 token consumption in late February 2026 and reached a full-platform majority around April 2026 on OpenRouter's rankings. Tokens are real work rather than press releases, and developers moved theirs.

Ecosystem and deployment

Adoption tells you most of what you need about ecosystem depth. Qwen's 200,000-plus derivative models on Hugging Face make it the single most-built-on open family, by Alibaba's own count. More derivatives means more fine-tunes and quantisations you can pull off the shelf instead of training yourself. It does not tell you whether the hosted Qwen3.8 is the right production endpoint.

[IMAGE: A developer's screen showing model weights downloading from a model hub, warm lighting, over-the-shoulder shot]

Each family owns a different deployment lane. Qwen3.6 gives you the most useful spread of open sizes and the gentlest fine-tuning path. DeepSeek V4 Flash is hard to beat for low-cost hosted reasoning, while Pro is for work that justifies more compute. Llama 4 still wins when the surrounding Western tooling and procurement comfort matter more than the licence difference.

All three have downloadable weights, but the hosted route is not symmetrical. DeepSeek runs its own V4 API. Qwen and Llama availability, regions and terms depend on the exact Alibaba or third-party service you choose. You can rent inference to get moving, then bring the same open checkpoint in-house when volume or data residency demands it. Verify that it really is the same checkpoint, not a newer hosted-only model such as Qwen3.8-Max.

Which one for which job

Match the model to the workload, not to the headline. The March 2026 capability-gap figure predates every Chinese release compared below, so each recommendation is about current fit, cost and licence rather than a stale country-level percentage.

Coding assistant

Start with DeepSeek V4 Flash for high-volume coding calls, then test Qwen3.6-27B if local control or a smaller open checkpoint matters. If the task runs for hours across a large repository, this three-family article is no longer the whole shortlist: Kimi K3 and GLM-5.2 now deserve the same evaluation.

RAG and back-office

Reach for Qwen3.6. Retrieval and internal-workflow jobs reward breadth, multilingual handling and clean fine-tuning, and Apache-2.0 keeps the first legal review straightforward. This is the lane where Qwen's variant library earns its keep.

Customer-facing

This one depends on your risk appetite. Llama 4 is the conservative pick if your stakeholders want a Western-origin model and you can live with the licence clauses. If governance signs off on a Chinese-origin model, Qwen3.6 gives you a clean open route and broad multilingual coverage. Either way, test the content behaviour your users will actually hit.

On-prem and regulated

For air-gapped or regulated builds, licence, model size and data control decide it. Qwen3.6 is the more manageable self-host route. DeepSeek V4 is MIT-licensed but much larger. Both can run without sending prompts to China, provided you actually deploy the weights inside your perimeter. Clear the governance review first; my piece on whether DeepSeek is safe for enterprise walks through that diligence.

FAQ

Is Qwen better than Llama?

For most open build decisions I'd start with Qwen3.6 because it offers manageable sizes, Apache-2.0 and a large derivative ecosystem. Llama 4 wins when Western-origin procurement comfort and existing Meta tooling outweigh the licence difference. Qwen3.8 is the stronger raw model but means a custom licence and serious hardware; Qwen3.8-Max is the hosted route, not the self-host answer.

Is DeepSeek better than Qwen?

For a cheap hosted reasoning endpoint, DeepSeek V4 Flash is the clearer choice. For open sizes, fine-tuning and ecosystem breadth, Qwen3.6 is easier to build around. Qwen3.8-Max may change the hosted capability comparison, and the open 3.8 weights earn a look once the licence and hardware maths work for you.

Are Chinese open models safe to build on?

They can be, with diligence. Self-hosting the weights means your data need not leave your infrastructure, which removes the China-hosted API route from the risk assessment. You still need to assess content behaviour, provenance, licence and supply-chain security, which I cover in my enterprise safety guide.

Which open model is cheapest to run?

On a first-party API, DeepSeek V4 Flash is the price leader in this comparison at $0.14 per million cache-miss input tokens and $0.28 per million output tokens. For self-hosting, a smaller Qwen3.6 checkpoint may cost less overall because DeepSeek V4 Flash still has 284B total parameters. Token price and infrastructure price answer different questions.

The bottom line

The current shorthand is versioned. Qwen3.6 for a broad Apache-2.0 open line. DeepSeek V4 Flash for cheap hosted reasoning, with Pro when the harder work earns its higher cost. Llama 4 when Meta's ecosystem and Western-origin procurement comfort matter most. Qwen3.8's weights are out now, under a custom licence; Qwen3.8-Max remains the hosted endpoint. Read the exact model card, price the actual deployment and test the task you need done.

If you're weighing this up for a real product, the governance questions come first and my piece on whether DeepSeek is safe for enterprise takes that diligence end to end. Tell me your use case and I'll point you at the right starting model.