The new open model race has shifted power in AI

For a long time, "open model" seemed synonymous with a lesser alternative to frontier labs. In 2026, this image became outdated. MiniMax M3, Qwen3.6, Kimi K3, and DeepSeek V4 compete in the area that matters most to companies: long context, reasoning, programming, multimodality, agents, and inference costs.

The change is not just technical. When weights can be downloaded, evaluated, and executed on your own infrastructure, the buyer gains deployment, privacy, and negotiation options. But openness is not a binary word. There are permissive licenses, custom licenses, weights available without training data, and models that require commercial authorization under certain conditions.

This article separates capability, license, and infrastructure with a cutoff date of 4 August 2026. The thesis is simple: open models already redistribute power in artificial intelligence, but alone do not democratize compute, transparency, or responsibility.

Comparative overview as of 4 August 2026

Numbers come from model cards and official repositories. They do not make the families directly comparable: each provider measures tasks, hardware, and settings differently. What the table allows comparing is the project's direction.

MiniMax M3: long context stops being just an interface promise

The MiniMax M3 combines a mixture of experts architecture with about 428 billion total parameters and 23 billion activated per token. The model is natively multimodal and accepts text, image, and video. Its declared window reaches 1 million tokens.

The most interesting innovation is the MiniMax Sparse Attention, a way to reduce attention cost on very long sequences. The company claims gains of up to nine times in initial filling and fifteen times in generation, plus significant compute reduction per token on 1 million tokens. These are vendor results, not a universal performance seal, but they address a real bottleneck: long context is only useful when latency and memory allow its use.

The M3 also offers three reasoning modes, from direct to deeper. In practice, this tries to balance two often-competing needs: quick answers for simple tasks and more inference budget for planning, programming, and agents.

Openness, however, requires precision. The M3 license allows many uses and weight distribution but imposes commercial attribution, prior authorization for organizations above a certain size, and usage restrictions. Therefore, the correct description is weight model available under a custom and restrictive license, not open source in the Open Source Initiative sense.

Qwen3.6: active efficiency and multimodality in the same design

The Qwen3.6-35B-A3B model is the first open-weight model of the Qwen3.6 generation. It has 35 billion parameters but activates approximately 3 billion per token. This reduces inference cost without forcing the system to give up a large expert base.

The hybrid design combines traditional attention with Gated DeltaNet plus a vision layer. The native context is 262.144 tokens and can be extended to nearly 1 million. The family also preserves reasoning states between steps, particularly useful in agent programming: an agent can switch between reading, tool execution, and correction without “forgetting” the entire operational thought process.

Among the four families, Qwen offers the simplest licensing situation for adoption: the analyzed model uses Apache 2.0, a permissive license recognized by legal and engineering teams. This does not mean the entire system is automatically “open source AI” by OSI definition because training data, its provenance, and the full reproduction recipe have not been published. The responsible formulation is open weights with a permissive license.

The official Qwen announcement on X also makes the ambition clear: multimodality, code, and agents cease to be separate products and become a deployment family.

Kimi K3: border-scale enters the world of available weights

The Kimi K3 takes the scale argument to the extreme: 2,8 trillion total parameters, 104 billion active, and 1.048.576 tokens of context. The Moonshot combines Kimi Delta Attention, Gated Multi-Head Latent Attention, a residual connection technique called AttnRes, and its own visual encoder.

The goal is not only to answer long questions. The model was designed for extended horizon work: understanding codebases, maintaining state when using tools, analyzing images, and completing tasks requiring many steps. The company claims scaling efficiency 2,5 times better than the prior generation. Again, this is a vendor measurement; editorial merit is in noting that the architecture tries to make a gigantic model viable without fully activating its size per token.

The official announcement on X, in 16 July, anticipated the release of weights, which later appeared in the official repository and model card. The Kimi K3 license is custom. It grants broad rights but should not be confused with Apache, MIT, or other OSI-approved licenses. Here, the safest category is open weights under a proprietary license.

K3 also shows an important limit of the democratization narrative. Downloading weights does not mean being able to serve them with good latency. Models at this scale still require clusters, high-performance memory, and specialized engineering.

DeepSeek V4: compression to transform 1 million tokens into product

The DeepSeek V4 Pro has 1,6 trillion total parameters and 49 billion active. The V4 Flash version reduces the scale to 284 billion total and 13 billion active. Both offer 1 million tokens of context and an MIT license.

The main novelty is the combination of sparse and heavily compressed attention. According to the model card, at 1 million tokens the V4 Pro uses 27% of inference operations per token and 10% of key and value cache compared to the V3.2. This directly addresses the cost that usually ruins long context applications: it is not enough to fit; reading, retrieval, and generation must occur without exploding memory and cost.

The Pro and Flash versions also translate a product decision. The first seeks greater capacity; the second offers a lighter path for frequent tasks. The official DeepSeek announcement on X, in 24 April, introduced both variants and the context window.

MIT is a permissive license, but the same caveat applies: open weights and inference artifacts alone do not reconstruct the entire training process. DeepSeek V4 is a permissively licensed open-weight model, with relevant transparency in the report and model cards but not with the totality of data and decisions needed for full reproduction.

Open source, open weights, and source-available are not synonyms

The Open Source AI Definition by OSI requires freedom to use, study, modify, and share an AI system. For “study and modify” to be real, final parameters are not enough: code and sufficiently detailed data information are necessary to understand and recreate the system.

OSI itself uses open weights to describe models whose parameters can be obtained but whose training is not fully open. It is a valuable category. It allows behavioral auditing, adaptation, local execution, and competition. It just doesn’t deliver the same transparency as a fully reproducible project.

A useful taxonomy for buyers is:

• Open source AI: freedoms to use, study, modify, and share, with components and information required by the OSI definition.
• Open weights: weights available for download and adaptation; code, data, and training recipe may remain partial.
• Source-available: code or weights can be seen, but the license preserves restrictions incompatible with open source.
• Permissive license: MIT and Apache 2.0 reduce friction of use and redistribution, though they do not guarantee that all training ingredients are open.
• Custom license: rights and duties must be read case by case; the “open” label does not replace this analysis.

This distinction should be integrated into the same purchasing process as cost, security, and retention. Our guide on how to choose an AI vendor without buying risk explains why contractual documentation is part of the architecture.

NVIDIA: accelerator of openness and compute center of gravity

NVIDIA contributes to this wave on three layers. First, it provides accelerators used to train and serve many of these models. Second, it offers libraries, kernels, and tools such as NeMo, TensorRT-LLM, and NIM, which turn weights into inference services. Third, it publishes its own open families, data, and tools.

In January, the company gathered open models, data, and tools in the Nemotron and Cosmos lines. In March, it announced the Nemotron Coalition, with partner labs, data, and cloud compute. The deployment catalog already includes, for example, MiniMax M3 and DeepSeek V4 Pro.

This is concrete contribution: public optimizations, tuning recipes, and distribution reduce the time from “weight available” to “executable system.” But there is a paradox. The more the open ecosystem depends on CUDA, cutting-edge GPUs, and clouds with these accelerators, the more software openness may reinforce infrastructure concentration.

In other words, open weights does not mean open compute. Alternative hardware and independent runtimes expand competition, but compatibility, memory availability, and tool maturity still weigh heavily. For a company, this layer needs to be part of portability design, just as we do when discussing AI in production after the pilot.

Regulation in the US: what is law, proposal, order, and export rule

In 4 August of 2026, the United States has no single comprehensive federal law regulating all open-weight models. The scenario is fragmented and must be described by the nature of each instrument.

• The America’s AI Action Plan is a public policy plan, not a law passed by Congress. It recommends encouraging open source and open-weight AI.
• The National AI Legislative Framework is a White House proposal to guide legislation. It is not, by itself, an existing federal statute.
• The NSPM-11 is a presidential memorandum aimed at the national security apparatus. It guides federal agencies; it is not equivalent to a general law for companies.
• The Bureau of Industry and Security rules are effective export controls. The January 2026 review started case-by-case analysis of some advanced chip exports to China under conditions. Parts of the Export Administration Regulations also deal with model technology and weights in specific categories.
• The NTIA report on open model weights recommended monitoring and evidence evaluation. It is a policy report, not a binding prohibition.

The debate thus crosses four conflicting goals: innovation, competition, security, and control over strategic capabilities. Broadly restricting weights may protect against some risks but also concentrate power in the few labs able to operate closed models. Releasing everything without evaluation may increase malicious uses and hinder accountability. Chip controls add another layer: they limit training and inference capacity even when software circulates.

This overview is informative and does not constitute legal advice. Companies deploying models should assess licenses, sector, export destination, data processed, and jurisdiction with specialized support.

What this race changes for companies

The question is not “which model won?”. The answer changes by task, hardware, and week. The most enduring question is: how much of the operation remains under control when the model changes?

There are five practical decisions:

1. Test the task, not just the benchmark. Compare completion rate, quality, latency, and cost in real flows.
2. Treat license as a technical requirement. A change in use, billing, or distribution can alter obligations.
3. Separate the operating model. Tools, memory, observability, and human supervision need to survive engine swaps. See the logic of agentic engineering.
4. Measure full cost. Free weights may require expensive hardware, optimization, security, and staff.
5. Plan portability. Avoid invisible dependencies on a format, a GPU, a library, or a cloud.

Long-context models also increase the temptation to put everything in the prompt. This can create exposure and uncontrolled memory. The article on AI memory as an operational risk shows how to separate useful history from indiscriminate accumulation.

The conclusion: openness has become a competitive advantage — with layers

MiniMax M3, Qwen3.6, Kimi K3 and DeepSeek V4 prove that available weights already compete in context, agents, code, and multimodality. The novelty is not in a single winner, but in the speed with which frontier techniques enter models companies can inspect, adapt, and run outside an exclusive interface.

At the same time, the map remains uneven. Licenses vary. Training data remains opaque. Compute is concentrated. Open tools can strengthen the same supplier dominating accelerators. And U.S. regulation advances through plans, executive orders, export controls, and proposals, not a single federal rule.

For leaders, the best response is not to choose an ideological camp. It is to build an operation able to compare models, swap components, and keep control over data, cost and responsibility. If you want to map where this portability creates value in your company, schedule an assessment with XMACNA.

Frequently asked questions

Are open weights the same as open source?

No. Open weights mean the final parameters can be obtained. Open source AI, according to OSI, requires freedoms and additional information that allow effective use, study, modification, and sharing of the system.

Which of the four families has the most permissive license?

In the analyzed sample, Qwen3.6-35B-A3B uses Apache 2.0 and DeepSeek V4 uses MIT. MiniMax M3 and Kimi K3 use custom licenses. Still, a permissive weight license does not mean the data and the full training process are open.

Does an open-weight model always cost less?

No. It may eliminate or reduce cost per call but add expenses with GPUs, inference engineering, security, availability, and support. The calculation must consider the total cost per completed task.

Does NVIDIA help or concentrate the open model market?

Both can be true. The company publishes tools, optimizations, models, and data that speed adoption. At the same time, reliance on CUDA and top GPUs concentrates a relevant part of infrastructure.

Have the United States banned open-weight models?

There is no general federal ban in effect as of the cutoff date. There are export controls, executive orders, sector policies, legislative proposals, and state rules. The effect depends on model, hardware, user, destination, and sector.