The Sequence
8/10 signal
The Sequence Radar #897: Last Week in AI: China, Compression and the Open-Model Race
modelsresearchsafety
Summary
This week in AI saw a shift from a singular focus on model scale to a multi-front competition over distribution, adaptability, and governance. Key releases included Thinking Machines' open-weight 975B-parameter MoE model Inkling, Moonshot AI's 2.8T-parameter Kimi K3, PrismML's hyper-compressed Bonsai 27B for mobile devices, and OpenAI's automated red-teaming model, GPT-Red.
Context
For years, AI progress has been framed as a straightforward race to build larger models with higher benchmark scores, dominated by a few well-funded labs. This narrative has centered on a centralized, cloud-based model of AI development and deployment. This week's developments challenge that paradigm by introducing competing visions for AI's future. Instead of just one axis of competition (scale), new fronts have opened up around open-weights vs. proprietary systems, cloud vs. edge deployment, and the scaling laws of safety and robustness. These releases represent different answers to the question of who will possess, adapt, and govern advanced AI, moving the conversation from pure capability to include accessibility, portability, and control.
Details
A Multi-Front Model Race:
The week's releases showcased competing strategies for AI development, moving beyond raw scale to focus on openness, compression, and safety.
| Model | Lab | Key Specifications | Strategic Focus |
|---|---|---|---|
| Inkling | Thinking Machines Lab | 975B parameters (MoE), 41B active. Native text, image, audio. 1M token context. Open weights, Apache 2.0 license. | Providing models that users can shape and customize, emphasizing controllable reasoning effort over leaderboard scores. |
| Kimi K3 | Moonshot AI | 2.8T parameters, 16 of 896 experts activated. Vision support, 1M token context. | Long-horizon coding and knowledge work. Billed as the first open model in the 3T-parameter class (full weights pending). |
| Bonsai 27B | PrismML | Ternary model (5.9GB) and one-bit variant (3.9GB). | Extreme compression to bring inference to edge devices, claiming to fit within a modern smartphone's usable memory while retaining baseline performance. |
Scaling Safety with GPT-Red:
- OpenAI introduced GPT-Red, an internal, automated red-teaming system designed to find vulnerabilities in other models.
- Mechanism: It is trained through self-play reinforcement learning to attack models, observe defenses, and create progressively stronger prompt injections.
- Performance: In tests, GPT-Red successfully compromised GPT-5.1 in 84% of scenarios, compared to a 13% success rate for human red-teamers.
- Impact: The attacks generated by GPT-Red were used to make GPT-5.6 significantly more robust against malicious instructions. This suggests a new scaling law for safety, where testing systems can evolve in capability alongside the models they are designed to secure.
Geopolitics and Industry News:
- At Shanghai’s World AI Conference, Xi Jinping positioned open-source AI as a global public good and promoted a new international AI cooperation organization, framing openness as a strategic lever for Chinese influence.
- Reuters reported China’s approval of Apple Intelligence, which will integrate Alibaba’s Qwen models. Baidu is also working with Apple for Chinese users.
- Demis Hassabis proposed a FINRA-style independent standards body to review frontier models up to 30 days before release.
- Reflection AI signed a $1 billion+ compute deal with Nebius for Nvidia GB300 chips, running through 2029.
- Bloomberg reported that Google's Gemini 3.5 Pro is months behind schedule due to missing internal goals, particularly in coding.
- SK Hynix raised $26.5 billion in the largest-ever US listing by a foreign company, selling 177.9 million ADRs at $149 each.
- Walden Robotics, a Toyota Research Institute spinout, launched with ~$300 million in seed funding at a $1.1 billion valuation.
What's new
The primary novelty is the fracturing of the AI competition narrative from a monolithic race for scale into a multi-dimensional contest. This is demonstrated by the simultaneous release of models prioritizing different strategic goals: user adaptability (Inkling), open-source scale (Kimi K3), extreme compression for the edge (Bonsai 27B), and scalable, automated safety systems (GPT-Red). This shift is coupled with an explicit geopolitical dimension, with China's leadership championing open-source AI as a tool for global influence.
Limitations
The article is a summary of recent announcements, and many of the claims from the model creators are not yet independently verified. For instance, the article notes that Kimi K3's full weights are not yet released and that independent testing is needed to confirm the performance retention of PrismML's compressed Bonsai 27B model.
The take
This week marks a clear inflection point where the definition of 'winning' in AI is becoming more complex. The race is no longer just about building the largest model but about controlling the ecosystem. The simultaneous pushes toward massive open models, hyper-efficient edge models, and scalable safety systems signal a maturing field where different strategic niches are being carved out. The most significant undercurrent is the weaponization of 'openness' as a geopolitical tool by China, aiming to build dependencies through standards and ecosystems rather than just hardware. The future of AI is not a single technological frontier but an emerging world order with competing power blocs, standards, and philosophies of control.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.