# Open weights gain at the bottom as the frontier closes at the top
> DeepSeek and Qwen keep open models near-frontier and cheap; Meta and Alibaba keep their best tiers proprietary

**Meta:** type: story · date: 2026-06-01 · heads: 他们没说的, 长远之局 · 6 takes · 3 lenses · 5 regions

## Summary

The open-vs-closed line is splitting by tier in 2026. Near-frontier open weights are advancing and cheap: [DeepSeek V4](/zh/n/deepseek-v4-open-weights-2026) (1.6T MoE, open on Hugging Face) and [Alibaba's Apache-2.0 Qwen 3.6](/zh/n/qwen-37-agents-open-vs-closed-2026) keep open models close to the frontier. But the very top is closing: Alibaba keeps Qwen 3.7-Max/Plus API-only, and [Meta abandoned open Llama](/zh/n/meta-superintelligence-talent-raid-2026) for closed Muse models. [Mistral](/zh/entity/mistral) holds an open-weight European position. US closed labs ([Openai](/zh/entity/openai), [Anthropic](/zh/entity/anthropic)) face the new wrinkle that [export controls](/zh/n/fable5-ai-export-controls) gate their top models by nationality while open Chinese weights face no equivalent restriction, an asymmetry reshaping the non-US market.

## By the numbers

- 1.6T, DeepSeek V4-Pro params (open weights).
- Apache 2.0, licence on Qwen 3.6 open line.
- Closed, Qwen 3.7-Max/Plus and Meta's Muse Spark.
- 0, export restrictions on open Chinese weights vs. gated US models.

## Why it matters

The competitive map is no longer "open vs closed" but "open-and-cheap at the near-frontier vs closed-and-gated at the top." Open Chinese weights gain structural advantage in markets where US models are restricted, while Western leaders monetise closed frontiers, bifurcating the global stack by both licence and nationality.

## What to watch

- Whether any lab open-weights a true frontier model in 2026.
- US policy toward open-weight Chinese models.
- Enterprise share shifting to open weights on cost grounds.

## Regional takes (batched by bias / lens)

### unlabelled
- **Hugging Face (DeepSeek-V4-Pro card)** (Global, en) — Open-weight model card for DeepSeek V4-Pro (1.6T MoE, 49B active, 1M context), the artefact at the centre of the open-frontier case: downloadable, near-frontier, far cheaper per token.
  Source: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
- **QwenLM (GitHub, Apache 2.0 line)** (China, en) — Alibaba's open Qwen 3.6 line under Apache 2.0, the open lower tier under the proprietary Qwen 3.7 flagships, illustrating the split-tier strategy.
  Source: https://github.com/QwenLM/Qwen
- **AI/ML API Blog** (Global, en) — 
  Source: https://aimlapi.com/blog/qwen-3-6-series-alibabas-open-source-llm-revolution-in-2026
- **Open Source For You** (India, en) — 
  Source: https://www.opensourceforu.com/2026/01/alibaba-launches-open-source-qwen-image-2512-as-a-serious-alternative-to-googles-image-ai/

### geopolitics / strategy
- **CNBC** (United States, en) — Frames China's open-weight push as a competitive lever that bypasses US closed-model access gates and chip controls, even as Chinese majors keep their top tiers proprietary for cloud revenue.
  > "China's open-source AI release intensifies the race as US labs close access to their most capable models."
  Source: https://www.cnbc.com/2026/04/24/deepseek-v4-llm-preview-open-source-ai-competition-china.html

### industry / Western pivot
- **The Next Web** (Netherlands, en) — Documents Meta abandoning the open-weight Llama tradition with closed Muse models, the clearest Western signal that the frontier is closing at the top even as open models advance below it.
  > "The IP these expensive hires produce will not be shared freely, Muse Spark is closed-source."
  Source: https://thenextweb.com/news/meta-thinking-machines-lab-talent-raid

## Across the graph
- Related: [[deepseek-v4-open-weights-2026]], [[qwen-37-agents-open-vs-closed-2026]], [[meta-superintelligence-talent-raid-2026]], [[fable5-ai-export-controls]]
- Entities: Deepseek, Alibaba Qwen, Meta AI, Mistral, Openai, Anthropic, China

---
Canonical: https://rbtfl.xyz/zh/n/open-vs-closed-frontier-2026