LotookLotookRSS feed
AI Insights · 5 min read

Open-Weight AI in 2026: What DeepSeek, Qwen, and Meta's Releases Mean for You

DeepSeek, Alibaba's Qwen, and Meta all shipped major open-weight models this year. What changed, why 'open' does not always mean open, and who should care.

By Chunyu Cai

Illustration of layered model blocks rising out of an open container, surrounded by network nodes

Most AI headlines are about closed models: the ones you can only use through a company’s app or API. But some of the most consequential releases this year were models whose weights anyone can download, run on their own hardware, and modify.

This fall, the open-weight side of AI got noticeably more crowded. Here’s what was released, what “open” actually means in each case, and who should care.

Open weights, briefly

A model’s weights are the billions of numbers it learned during training. When a company releases them, you can run the model yourself instead of sending your data to someone else’s servers. You can also fine-tune it on your own data.

That’s different from “open source” in the traditional software sense. Many open-weight models don’t publish their training data or full training code, and many come with licenses that restrict how you can use them. I’ll come back to that, because it matters more than most coverage suggests.

This year’s major releases

DeepSeek

The Chinese lab DeepSeek has kept up a steady cadence this year. Its V4 family arrived on Hugging Face in Pro and Flash versions in the spring, followed by updated checkpoints over the summer, an experimental vision model at the start of September, and a V4.1 Flash release in mid-October. DeepSeek’s frequent, incremental updates have made it one of the most closely watched open-weight labs.

Alibaba’s Qwen

On August 14, Alibaba’s Qwen team released open weights for two Qwen3.8 models. The smaller one, Qwen3.8-27B, is a 27-billion-parameter multimodal model that handles text, images, and video, with a native context window of about 262,000 tokens that the team says can be extended to one million. Qwen says it outperforms the larger Qwen3.7-Plus on coding and office tasks. It’s released under the permissive Apache 2.0 license.

The second, a much larger model called Qwen3.8-2.4T-A95B, is aimed at the top end. According to reporting by eWeek, it comes with a custom license that adds limits for some large commercial users.

Meta

Meta’s open-weight story took a turn. Llama 4, released in April 2025, was the last major release under the Llama name. This year Meta moved to a new model family called Muse, and its first Muse model shipped in April as a closed, API-only product.

On October 15, Meta released Muse Spark Open, a large mixture-of-experts model with open weights. Meta’s chief AI officer, Alexandr Wang, called it “the most capable US open-weight model to date,” and Meta says it rivals, and in some cases outperforms, similar models from leading Chinese developers such as Alibaba’s Qwen, Moonshot AI’s Kimi, and Z.ai’s GLM. Those are Meta’s own claims.

Meta says the model can run on two Mac Studios with Apple’s M5 Ultra chips, which tells you something: “runs locally” now means serious hardware at the top end. It’s also released under Meta’s own license, with commercial restrictions, rather than a standard open-source license.

What actually changed

Chinese labs set the pace

The most telling detail of the fall may be how Meta framed its own launch. Its pitch for Muse Spark Open was built around how it compares with Chinese open-weight models. For the past year, labs like DeepSeek, Alibaba’s Qwen team, Moonshot, and Z.ai have defined what a strong open-weight model looks like, and the largest US entrant is now positioning itself against them.

For some businesses, that matters for reasons beyond performance. Companies in regulated industries, or with policies about where their software comes from, may need to check the origin and license of any model before using it, even when it runs entirely on their own servers.

Small models got genuinely useful

The releases I find most practical aren’t the giants. They’re models like Qwen3.8-27B: small enough to run on a single high-end GPU when compressed, yet capable enough for real work like document processing, internal search, and coding help. For a lot of business tasks, that’s the sweet spot.

“Open” now covers a wide range

Compare the licenses: Apache 2.0 for Qwen’s smaller model, a custom license with limits for its larger one, and Meta’s own license with commercial restrictions for Muse Spark Open. All of them get called “open.” They aren’t equally open. Always read the license before you build a product on a model, especially if you plan to grow.

Who should care

You probably don’t need open weights if you’re an individual or a small team using AI for everyday writing, research, and analysis. A good closed model through a business plan is simpler and usually stronger on the hardest tasks.

Open weights are worth a serious look if:

  • Your data can’t leave your environment. Running a model on your own infrastructure keeps sensitive data in-house. Our guide on using AI at work without leaking sensitive data covers when that matters.
  • You run high volumes. At scale, hosting your own model for routine tasks can cost less than paying per token.
  • You need to customize. Fine-tuning on your own data is far easier when you control the weights.
  • You want to avoid lock-in. A model you host can’t be deprecated, repriced, or changed underneath you.

The trade-off is that you take on what a provider normally handles: hardware, updates, monitoring, and safety measures. That’s real work, and it’s why many teams end up using both: closed models for the hardest, lowest-volume work, and open-weight models for high-volume or sensitive tasks.

My take

The open-weight ecosystem is healthier than it was a year ago: more labs, more model sizes, and more competition. That’s good for everyone, including people who never download a model, because it puts pressure on closed-model pricing.

But I’d be careful with the word “open.” The most important question about an open-weight model isn’t its benchmark score. It’s whether the license lets you do what you’re planning to do, and whether you can actually run it on hardware you can afford.

For the closed-model side of this fall’s releases, see what the GPT-6, Gemini 4 Argon, and Claude 5.5 launches changed. And if you’re choosing a model for a specific job, our guide on how to choose an AI model walks through the trade-offs, including open versus closed.

Sources

Written by

Chunyu Cai

Chunyu Cai is the CEO of Lotook, LLC and an entrepreneur who writes about AI models, agents, and how they change everyday work. Chunyu holds a master's degree from Imperial College London.

X (@ChunyuCai)

All articles →
Illustration of many small points of light entering a funnel, with only a few bright ones emerging at the bottom
AI Insights · 5 min ·

Everyone Is Piloting AI Agents. Almost No One Has Scaled Them.

Most companies are experimenting with AI agents, but few run them at scale. Why the gap exists, why size matters, and what the companies getting it right do.

Illustration of a ladder whose lowest rungs are dissolving into particles while the upper rungs stay solid
AI Insights · 5 min ·

Why Entry-Level Jobs Are the Front Line of AI

New payroll research shows young workers in AI-exposed jobs falling behind their peers. What the data says, why entry-level roles are hit first, and what to do.

Illustration of three glowing model nodes rising along a wave, representing three new AI model releases
AI News · 5 min ·

GPT-6, Gemini 4 Argon, and Claude 5.5: What the Fall 2026 Model Wave Changed

OpenAI, Google, and Anthropic all shipped new flagship models within a month. Here is what each released, what actually changed, and what it means for you.