This spring, Alibaba dropped Qwen3.5-Omni – a powerful multimodal model. The flagship versions? API-only. The lightweight Light got open weights.
Meta did the same. Muse Spark, the successor to the open LLaMA family that defined the movement, is now proprietary. For developers who bet on open models, this is a warning shot. Will more follow?
The Gap Is Widening Again
The long-term fate of open models is about economics, not technology. Models are more powerful, more expensive to train, and more strategically valuable. Sharing the frontier means giving away your edge.
According to the Stanford AI Index 2026, the quality gap between closed and open models is growing again. It was just 0.5% in August 2024. By March 2026, it hit 3.4%. Today, the entire top ten on the Arena AI leaderboard are closed models.
Here’s the twist: on SWE-bench Pro, open models actually lead. GLM-5.1 beats Claude Opus 4.6. But on complex math reasoning? The gap persists. Closed labs are deliberately building their lead in agentic tasks, multimodality, and reasoning.
New Strategy: Openness as Marketing
The industry is pivoting. Mid-tier models get open weights to capture the ecosystem and stay visible. The most powerful ones stay proprietary.
Anthropic took it further. The new Claude Mythos is available only to select partners. Developers claim it autonomously finds zero-day vulnerabilities and turns them into working exploits. The strongest models are now competitive advantages no one wants to give away.
Transparency is fading. Of 95 significant models released in 2025, 80 shipped without training code. The Stanford Transparency Index dropped 17 points. Leading labs won’t even share parameter counts anymore. Irony: AI got this far precisely because knowledge spread freely.
The Economics No Longer Favor Openness
For companies, open source was always marketing, not charity. When Meta opened LLaMA in 2023, the logic was clear: attract developers, make your architecture the standard. It worked for DeepSeek and Mistral too.
But this strategy has a shelf life. Training costs are growing at 2.4x per year since 2016. Dario Amodei said advanced models could cost $1 billion per run — with $10 billion runs ahead.
At those numbers, giving away your main asset is a tough sell to investors. OpenAI projects losses through 2030. Anthropic doesn’t expect positive cash flow until 2028. Closed models now account for roughly 80% of global AI usage and 96% of revenue. The open model share in enterprise fell from 19% to 11%.
Which Open Models Are Still Strong?
Companies still releasing open models haven’t disappeared. Open source handles most real-world tasks just fine. But if you have the resources, building your own makes growing sense – you control quality, cost, and independence.
- DeepSeek V4. Showed that frontier-level quality and open weights aren’t mutually exclusive. After its release, that became the new standard everyone expects.
- Kimi K3. The newest and largest open model here — 2.8 trillion parameters, which Moonshot calls the first “open 3T-class” model. Its Agent Swarm system still runs up to 300 coordinated sub-agents across 4,000 consecutive steps, where most models lose coherence after 20–50. On SWE-bench Verified it leads all open models at 80.2%. Available natively in Cursor — a top AI coding tool.
- GigaChat 3.5 Ultra. The only model here from neither the US nor China. Trained from scratch on 10 languages including Russian, Kazakh, and Arabic. On par with Qwen and DeepSeek in math, code and agentic tasks. Its Kandinsky image/video models are also well-regarded — even used by the AI artist Botto, whose works sold at Sotheby’s.
- GLM-5.2. A partnership between Zhipu AI and Tsinghua University. Excels at long decision chains, now with a 1-million-token context window. Currently ranks first among all open models on the Artificial Analysis Intelligence Index. Also led all frontier models in honesty — it knows when to say “I don’t know,” reducing hallucinations.
- Gemma 4. Google keeps Gemini closed but keeps releasing strong open models under Gemma. Tops open leaderboards in coding. Multimodal capabilities are available even in compact versions that run on consumer hardware.
What This Means for Businesses
When open releases stop, you’re stuck: stay on the current version and accept the growing gap or switch models and risk breaking what works.
Fine-tuning on proprietary data embeds unique expertise into the model. With closed platforms, everyone gets the same base intelligence — differentiation happens only at the app layer.
Then there’s vendor lock-in. A closed provider can change pricing, update the model, or revise terms – and your product gets hit immediately. With open weights, you decide when and whether to update.
Disclaimer: GeekWire newsroom and editorial staff were not involved in the creation of this content..
