DeepSeek, Qwen & the Open-Weight Wave
For a long time the story of "good AI" was the story of a handful of closed, frontier models you could only rent through an API. That's no longer the whole story. A wave of open-weight models โ many from Chinese labs like DeepSeek and Alibaba's Qwen, alongside Meta's Llama and France's Mistral โ got good enough, cheap enough, and small enough that downloading and running your own capable model became a real option. This page is the durable map of that landscape: why it happened, what "open weight" actually means, the honest trade-offs versus closed frontier models, and how to actually use these models today.
- Understand WHY open-weight models became competitive โ capability-per-cost and open reasoning
- Pin down what 'open weight' really means โ vs open-source and vs open-data
- Recognize the archetypes: general chat, reasoning, coding, and small/efficient models
- Weigh the real trade-offs vs closed frontier โ capability gap on the hardest tasks, and owning your own infra and safety
- Know the two ways to actually use them: a hosted API, or self-hosting via Ollama
Why open-weight models suddenly got competitiveโ
Three things shifted at roughly the same time, and together they closed most of the everyday gap.
- Capability-per-cost collapsed. Better training recipes, data curation, and mixture-of-experts (MoE) architectures (where only a fraction of a large model's parameters activate per token) let labs ship models that punch far above their cost. For a large share of real tasks โ drafting, summarizing, extraction, classification, routine coding โ a good open model is now "good enough," at a fraction of frontier-API cost.
- Open reasoning models arrived. DeepSeek's reasoning line was a turning point: it showed that the "think step by step before answering" capability โ previously the moat of a few closed models โ could be trained and released with open weights. Once one lab proved it, the rest of the open ecosystem followed.
- Small models got genuinely usable. Sub-10B and even sub-1B models became coherent and useful, which means you can run something capable on a laptop or a cheap server. That changes the economics of privacy, latency, and scale.
The net effect: the gap to the closed frontier didn't vanish, but for most work it stopped being the deciding factor. The deciding factors became cost, control, and where your data lives.
What "open weight" actually meansโ
This is the single most-confused point in the whole space. "Open" is doing a lot of work in three different ways, and they are not the same thing.
| Term | What you actually get | What you usually DON'T get |
|---|---|---|
| Open weight | The trained model file (the weights). You can download, run, fine-tune, and self-host it. | The training data, and sometimes the full training code. Use may be governed by a custom license. |
| Open source (strict) | Weights plus training/inference code under an OSI-style license, with freedom to use, modify, and redistribute. | Still often not the training data. |
| Open data | The actual datasets the model was trained on, openly published. | Rare for frontier-scale models; most "open" models are open-weight, not open-data. |
The practical takeaway: most models people call "open source" are really open-weight. You can run them and adapt them, but you typically cannot reproduce them from scratch, and you must read the license. Licenses vary:
- Some ship under permissive licenses (Apache 2.0, MIT) that are clean for commercial use โ Qwen and Mistral models, for example, have historically used Apache 2.0, and DeepSeek has used MIT for major releases.
- Some ship under custom community licenses with conditions โ Meta's Llama license, for instance, has historically included a clause restricting the very largest companies (by monthly active users) from using it freely.
Always check the specific model card for its license before building on it. Neutral but important: a model being open-weight tells you that you can host it yourself; it does not tell you the legal terms, and it does not make the licensing or data-residency questions go away โ it relocates them to you.
The four archetypesโ
Open-weight families almost always come as a spread of variants, not a single model. Learn the four archetypes and you can read any lab's lineup at a glance.
A single family โ Qwen is the clearest example โ typically gives you the whole ladder: tiny models for edge and laptops, mid-size models for servers, a large flagship for hardest tasks, plus coding- and reasoning-specialized variants. You pick the rung that fits your hardware and task, not "the model."
The honest trade-offs vs closed frontierโ
Open-weight models are not a strict upgrade. They're a different deal. Here's the trade neutrally.
| Dimension | Open-weight (self-host or cheap host) | Closed frontier (hosted API) |
|---|---|---|
| Hardest-task capability | Strong and closing fast, but the closed frontier usually still leads on the very hardest reasoning, long-horizon agentic, and niche tasks | Typically the ceiling on the toughest tasks |
| Cost at scale | Can be dramatically cheaper, especially self-hosted or via budget hosts | Premium per-token; predictable but adds up |
| Privacy / data residency | You can keep data fully inside your environment | Data leaves your network to the provider |
| Control & customization | Full: fine-tune, quantize, pin a version, run offline | Limited to what the API exposes; versions can change under you |
| Who owns safety & ops | You do โ guardrails, abuse filtering, uptime, scaling, and patching are your job | The provider runs infra and ships baseline safety |
That last row is the one people underweight. When you self-host an open-weight model you also inherit the responsibilities a frontier lab normally absorbs for you: content safety and abuse mitigation, keeping the deployment patched, capacity and uptime, and evaluating each new model version yourself. "Open weight" buys you control, and control is a cost as much as a benefit. For cost intuition on the API side of this trade, see Token economy.
How to actually use an open modelโ
There are exactly two paths, and most teams use both at different stages.
- Hosted API โ many providers (including the labs themselves and third-party inference hosts) serve open-weight models behind an API, often OpenAI-compatible. You get open-model economics with zero infrastructure work. Best when you want low cost without running servers, but note your data still leaves your network.
- Self-host โ download the weights and run them in your environment. Maximum privacy and control, and the only option when data truly cannot leave your network. The easiest on-ramp is Ollama, which pulls and runs open models with a single command.
Try a model locally in four stepsโ
- Download it from ollama.com/download (macOS/Windows) or use the Linux install script. This gives you the ollama command and a local background service. Full walkthrough on the Run models locally page.
- Smaller models need less RAM/VRAM. A laptop comfortably runs small variants (a few GB); large flagships need a real GPU or server. Start small to confirm everything works, then scale up if the quality isn't enough.
- ollama run <model> downloads the weights on first use (several GB), then drops you into an interactive chat. First run is slow because it's pulling; later runs are instant from cache.
- Ollama exposes an OpenAI-compatible local endpoint, so you can point an existing OpenAI-style SDK at localhost instead of the cloud. Now the same code runs against a private, local model.
Pull and run an open model locally with Ollama
ollama run qwen3
That one command downloads a Qwen open-weight model and starts a local chat โ no API key, no data leaving your machine. Swap qwen3 for another family (for example a Llama, Mistral, or DeepSeek tag) to compare them on your own task. Picking the right model for a given job is its own skill: see Choosing a model.
Choosing between open and closed, in practiceโ
Don't pick by leaderboard. Pick by constraint:
- Data cannot leave your network โ self-hosted open-weight is the answer, full stop.
- High volume, cost-sensitive, "good enough" quality โ open-weight (hosted or self-hosted) usually wins on cost.
- You need the absolute best on the hardest tasks โ the closed frontier still tends to lead โ use it where it matters and route easier traffic to cheaper open models.
- You want to fine-tune, quantize, pin a version, or run offline โ only open weights give you that control.
The most durable move is to run a tiny evaluation on your own real data across one open and one closed candidate. Benchmarks describe someone else's task; your eval describes yours.
Check yourself
0/3- Open-weight models closed most of the everyday gap thanks to capability-per-cost, open reasoning models, and genuinely usable small models
- 'Open weight' = downloadable weights you can run and fine-tune โ NOT necessarily open-source code, open training data, or a permissive license; always read the model card
- Read any lineup by archetype: general chat, reasoning, coding, and small/efficient โ most families ship the whole size ladder
- The trade is real: open buys cost, privacy and control, but the closed frontier usually still leads on the hardest tasks and you inherit safety and infra
- Two ways to use them: a hosted (often OpenAI-compatible) API, or self-host via Ollama with one command โ then evaluate on YOUR data, not a leaderboard
Sources & further readingโ
- DeepSeek on Hugging Face (official organization)
- DeepSeek on GitHub
- Qwen on Hugging Face (Alibaba, official organization)
- Qwen on GitHub
- Meta Llama โ official site
- Llama models on Hugging Face (meta-llama)
- Mistral AI on Hugging Face
- Hugging Face โ Models hub
- Ollama model library
- Related on AILmanac: Kimi K2 for Claude users โ Moonshot's trillion-parameter open agent, in depth
- Related on AILmanac: GLM-5.2: Open-Weight Frontier Coding Model โ Z.ai's 753B MoE that closes to within a point of Claude Opus 4.8 on long-horizon coding