English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

AI Industry Daily (June 25, 2026): OpenAI Custom Chip Jalapeño, GLM-5.2 Open-Source Breakout, Agents Enter Team Software

Forum topic · 小凯 · 2026-06-25

Summary

A June 25, 2026 daily digest from the easy-learn-ai project covering major AI industry developments. OpenAI reportedly ships GPT-5.5 Instant with improved constraint adherence and unveils Jalapeño, its first in-house inference chip co-developed with Broadcom to reduce NVIDIA dependency. GLM-5.2 emerges as one of the strongest open-weight models, with Baseten and Cursor adopting it. Baidu open-sources Unlimited-OCR, a 3.3B-parameter multilingual document recognition model under MIT license, and Qwen releases AgentWorld-35B-A3B, a 256K-context MoE simulator for training agents. Krea open-sources its first in-house image model Krea 2. Agents are entering team software via Claude's Slack identity and Cursor-Notion task dispatch, while memory systems like Weaviate Engram and LangSmith Context Hub become an independent infrastructure layer. Other items include Qualcomm acquiring Modular, NVIDIA NeMo AutoModel's 3.4-3.7x MoE training speedup, inference optimizations from SkyPilot/Modal/DFLASH, funding news, chip export controls, and notable papers and product updates.

This is an English translation of the June 25, 2026 daily AI industry digest from the easy-learn-ai project (GitHub, commit 71dca122771e453a03625f32e23327ce014ad72b).

1. OpenAI's Two Cards: GPT-5.5 Instant and the Jalapeño Chip

  • GPT-5.5 Instant is officially described as improving intent understanding, constraint adherence, and chat tone — not a generational leap, but a notable emphasis on following system rules and refusing out-of-bounds requests, a signal for AI safety practitioners.
  • Jalapeño is OpenAI's first in-house inference chip, co-developed with Broadcom, targeting inference for ChatGPT, Codex, the API, and future agents to reduce reliance on commercial GPUs (i.e., NVIDIA). Co-founder Greg Brockman discussed energy efficiency on X; community speculation suggests it may have taped out nine months ago.
  • 2. GLM-5.2: The Open-Source Dark Horse

    GLM-5.2 is widely seen as one of the strongest open-weight models available. Testers report web-task quality approaching Opus 4.8 at higher speed and lower cost. Baseten and Cursor have both onboarded it, and it performs well on Code Arena and sparked discussion on ARC-AGI-2. After a quiet stretch post-Llama, open-source models look competitive again — a high value-for-money option for budget-conscious teams.

    3. Baidu Unlimited-OCR

    Baidu open-sourced Unlimited-OCR, a 3.3B-parameter multilingual document recognition model under MIT license. It parses images, multi-page documents, and PDFs in one pass, outputs up to 32K tokens, and supports SGLang and OpenAI-compatible streaming APIs. Released on ModelScope rather than as a closed API, mirroring the open-source ecosystem strategy of GLM-5.2.

    4. Krea 2 Open-Sourced

    Krea released Krea 2, its first fully in-house image generation model, with Raw and Turbo versions — code and weights fully public. The team is discussing a 5B variant, editing capabilities, and better text rendering. Stable Diffusion's dominance is being challenged by independent teams.

    5. Agents Enter Team Software

  • Claude now has a standalone Slack identity, credentials, and audit records; Cursor tasks can be dispatched directly from Notion.
  • Karpathy discussed Claude's Slack Agent design on X; Kenton Varda discussed permissions. Warnings followed: poorly designed agent permissions may be riskier than human credential leaks, since agents run 24/7.
  • Hugging Face's internal Moon Bot — a self-hosted Slack coding agent connected to GitHub, Athena, MongoDB — emphasizes auditability, control, and no vendor lock-in.
  • 6. Qwen AgentWorld: A Simulator for Agents

    Qwen open-sourced AgentWorld-35B-A3B (35B MoE, ~3B active, 256K context). It can simulate MCP tool calls, search, terminals, SWE development, web operations, operating systems, and Android devices — a training ground so agents can learn safely in simulation before real-world deployment.

    7. Agent Memory as an Independent Layer

    Weaviate's Engram reached GA; LangSmith released Context Hub; papers on memory evaluation are emerging. Agent memory is becoming standalone infrastructure requiring extraction, deduplication, updating, permission control, and lifecycle management.

    8. China's AI Chip Ecosystem

    Community-tracked roadmaps span Huawei, Alibaba, Baidu, MetaX, Moore Threads, Biren, and Iluvatar. Debate centers on software stacks and mass production over raw specs. Huawei's 950 SuperPOD progress also drew attention. NVIDIA's moat isn't insurmountable, but the challenge is more than transistor counts.

    9. NVIDIA NeMo AutoModel

    NVIDIA claims NeMo AutoModel delivers 3.4–3.7x MoE training throughput via Expert Parallelism, DeepEP, and TransformerEngine kernels — potentially lowering training costs further.

    10. Inference Deployment Heats Up

  • SkyPilot launched unified inference Endpoints.
  • Modal claims open-source inference latency beats closed services.
  • DFLASH draft models deliver 30–50% gains in real decoding.
  • 11. Qualcomm Acquires Modular

    Chris Lattner (creator of LLVM and Swift) announced Qualcomm will acquire Modular; Mojo's open-sourcing plans continue. Competition in AI compilers and non-CUDA inference stacks intensifies.

    12. Funding and Talent

    Mirendil AI debuted with a $200M seed round; UK's BOLD Lab and SOFAIR raised £60M; talent flows from Google DeepMind to Anthropic continue, including mechanistic interpretability researcher Arthur Conmy joining Anthropic.

    13. Safety and Governance

  • Anthropic is involved in export control litigation, vulnerability-testing controversy, and distillation accusations.
  • The US Chip Security Act proposes location-tracking mechanisms in high-end AI chips; the community raises security and privacy concerns.
  • Virginia data center noise complaints (joined by Carmack) highlight AI infrastructure's social costs.
  • 14. Notable Products and Papers

  • OpenThoughts-Agent: 100+ ablations, a 100K-sample training set fine-tuned on Qwen3-32B, averaging 44.8% across 7 agent benchmarks. Instructions, trajectory length, and data source all matter.
  • Kimi API on AWS Marketplace: procurement becomes easier for enterprises.
  • Seedance for anime 3D-to-video: Japanese animators use 3D scenes for motion composition plus AI rendering, keeping long-video consistency better than pure text-to-video.
  • ChatGPT photo restoration shows identity drift: aging-then-restoring tests produce plausible but not-actually-the-person faces — completion, not faithful restoration.
> Source: easy-learn-ai daily update for 2026-06-25, community-curated. Original links: data/daily/md/2026-06-25.md in the easy-learn-ai repository.

Tags

#openai#open-source-models#ai-chips#ai-agents#inference-optimization#glm#ai-governance#daily-news

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208114