[AINews] OpenAI shuts off Cursor
OpenAI terminated Cursor's API access citing "experience with Elon Musk's companies violating contracts," marking a significant escalation in the ongoing rivalry between the two camps Z.ai released GLM-5.3 as open-weight models targeting agentic coding and cyber defense, with vLLM confirming day-0 support for 744B total / 40B active parameters, 1M context, and 128K max output Tencent's Hy4-preview emerges as a top-tier open MoE (770B total / 49B active, 1M context) with a novel serving design fe
Analysis
TL;DR
- OpenAI terminated Cursor's API access citing "experience with Elon Musk's companies violating contracts," marking a significant escalation in the ongoing rivalry between the two camps
- Z.ai released GLM-5.3 as open-weight models targeting agentic coding and cyber defense, with vLLM confirming day-0 support for 744B total / 40B active parameters, 1M context, and 128K max output
- Tencent's Hy4-preview emerges as a top-tier open MoE (770B total / 49B active, 1M context) with a novel serving design featuring 256 routed experts + 1 shared, only 21/78 layers computing sparse indices, plus a 10B MTP draft depth-3 layer
- Qwen3.8-Flash offers a cheap long-context MoE at 125B total / 6B active with 1M context and multimodality, priced at ~$0.15/1M input and $0.47/1M output, though FP8 quantization stability issues were reported
- vLLM published a benchmark-driven comparison of speculative decoding methods (MTP, EAGLE-3, DFlash, DSpark) across multiple model families on AMD hardware, concluding there is no universal winner and teams should treat it as a tuning surface
Why It Matters
The OpenAI-Cursor split signals a hardening of competitive boundaries in the AI coding tool ecosystem, with both OpenAI and xAI/Grok now positioning themselves as serious coding alternatives to Claude. Simultaneously, the open-weight frontier is accelerating rapidly with GLM-5.3, Hy4-preview, and Qwen3.8-Flash providing deployable alternatives to closed models, while vLLM's guidance helps practitioners make informed infrastructure decisions about speculative decoding.
Technical Details
- GLM-5.3: Open-weight family positioned for agentic coding and cyber defense; vLLM supports 744B total / 40B active, 1M context, 128K max output; UnslothAI claims a 239GB 2-bit variant retaining ~81% accuracy from the original 1.51TB; GLM-5.3-Flash achieves 270 tok/s at 1/10 the cost with 10% higher quality than GLM-5.2 on OfficeQA Pro v2
- Tencent Hy4-preview: 770B total / 49B active MoE with 1M context, ranked ~#5 on Code Arena: WebDev (+115 pts over Hy3), leads on SWE-bench Pro; serving architecture uses 256 routed experts + 1 shared expert, with only 21/78 layers computing their own sparse index while others reuse it, plus an embedded 10B MTP layer with draft depth 3
- Qwen3.8-Flash: 125B total / 6B active MoE with 1M context and multimodality; approximately 20x cheaper and ~2x faster than Qwen3.8 Max; pricing at $0.15/1M input and $0.47/1M output; FP8 quantization caused broken multi-turn tracking, resolved by switching KV cache to BF16
- vLLM Speculative Decoding: Benchmark comparison of MTP, EAGLE-3, DFlash, DSpark, and a fifth method across Gemma, Qwen, Kimi, and MiniMax on AMD MI300X/MI355X; no universal winner — optimal method depends on model family, workload, and speculation depth
Industry Insight
- The OpenAI-Cursor split demonstrates that API access is becoming a strategic weapon in the coding AI race; companies should diversify model dependencies and prepare for potential access disruptions as competitive dynamics intensify between OpenAI, Anthropic, and xAI ecosystems
- Open-weight frontier models are reaching parity with closed alternatives — the GLM-5.3, Hy4, and Qwen3.8-Flash releases show that self-hosted deployment is now viable for production workloads, reducing vendor lock-in and enabling cost optimization through quantization and custom serving infrastructure
- Speculative decoding should be treated as an ongoing optimization surface rather than a one-time configuration; teams investing in inference performance should benchmark multiple methods against their specific model-family and workload characteristics, especially as AMD GPU adoption grows
Disclaimer: The above content is generated by AI and is for reference only.