AI Industry Daily Roundup (June 25, 2026): OpenAI In-House Chips, GLM-5.2 Open-Source Surge, Agents Enter Team Software

Forum topic · 小凯 · 2026-06-25

Summary

A June 25, 2026 daily AI industry digest from the easy-learn-ai community covering major model, hardware, and infrastructure developments. OpenAI updated GPT-5.5 Instant with improved instruction and constraint following, and revealed Jalapeño, its first in-house inference chip co-developed with Broadcom to reduce reliance on NVIDIA. GLM-5.2 is widely regarded as one of the strongest open-weight models, approaching Opus 4.8-level web task quality with lower cost, and is already live on Baseten and Cursor. Baidu open-sourced Unlimited-OCR, a 3.3B-parameter multilingual document recognition model under MIT license. Krea released its first fully in-house image generation model Krea 2, and Qwen open-sourced AgentWorld-35B-A3B, a world simulator for agent training supporting 256K context. Additional items: agents gaining enterprise identities in Slack and Notion, agent memory emerging as an infrastructure layer (Weaviate Engram GA, LangSmith Context Hub), Qualcomm acquiring Modular, NVIDIA NeMo AutoModel MoE training speedups of 3.4–3.7x, and US chip location-tracking legislation discussions.

Source

Commit: 71dca122771e453a03625f32e23327ce014ad72b Project: easy-learn-ai (https://github.com/ConardLi/easy-learn-ai)

Key points

1. OpenAI's two big cards: GPT-5.5 Instant and in-house chip Jalapeño

GPT-5.5 Instant is described officially as improvements in intent understanding, constraint adherence, and chat tone — not a generation jump, but a notable signal that constraint-following (system rule compliance, refusal of out-of-bounds requests) was singled out.
Jalapeño is OpenAI's first in-house inference chip, co-developed with Broadcom. It targets inference for ChatGPT, Codex, API, and future agent workloads, reducing dependence on commercial GPUs (i.e., NVIDIA). Co-founder Greg Brockman discussed energy efficiency on X; the community speculates tape-out happened about 9 months ago.

2. GLM-5.2: the open-source dark horse

Widely considered one of the strongest open-weight models available. Testers report near Opus 4.8 web-task quality at higher speed and lower cost. Already live on Baseten and Cursor; strong Code Arena rankings and discussion around ARC-AGI-2.
After a quiet stretch following Llama, GLM-5.2 revives confidence in open-source competitiveness — a high value-for-money option for budget-conscious teams.

3. Baidu Unlimited-OCR

Open-sourced under MIT license: a 3.3B-parameter multilingual document recognition model supporting images, multi-page documents, and PDFs in one pass, up to 32K output tokens, with SGLang and OpenAI-compatible streaming interfaces. Released on ModelScope rather than as a closed API — signaling a community-oriented ecosystem strategy among Chinese LLM vendors.

4. Krea 2 open-sourced

Krea's first fully in-house image model, released with code and weights in Raw and Turbo variants; the team is debating a 5B version, editing capabilities, and better text rendering. Independent teams are now openly challenging Stable Diffusion's dominance.

5. Agents enter team software

Claude now has an independent Slack identity, credentials, and audit trail; Cursor tasks can be dispatched directly from Notion. Karpathy discussed Claude's Slack agent design; Kenton Varda discussed permissions. A warning circulating: 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 wired to GitHub, Athena, MongoDB — emphasizing auditability, control, and no vendor lock-in.

6. Qwen-AgentWorld: a world simulator for agents

AgentWorld-35B-A3B: 35B-parameter MoE with ~3B active parameters and 256K context. It can simulate MCP tool calls, search, terminals, SWE development, web browsing, OS, and Android devices — a training ground so agents can learn skills safely in simulation before real deployment.

7. Agent memory is becoming an independent system layer

Weaviate's Engram reached GA; LangSmith released Context Hub; papers discuss agent memory evaluation. Memory now requires extraction, deduplication, updating, permission control, and lifecycle management — essential for multi-turn dialogue and long-horizon tasks.

8. China's AI chip ecosystem heats up

Community-compiled roadmaps for Huawei, Alibaba, Baidu, MetaX, Moore Threads, Biren, and Iluvatar. Debate: hardware specs matter less than software stacks and mass production. Huawei's 950 SuperPOD progress also drew attention.

9. NVIDIA NeMo AutoModel

Claims 3.4–3.7x MoE training throughput gains via Expert Parallelism, DeepEP, and TransformerEngine kernels — potentially lowering training costs.

10. Inference deployment competition

SkyPilot launched unified inference Endpoints; Modal claims open-source inference latency beats closed services; DFLASH draft models deliver 30–50% real-world decoding speedups.

11. Qualcomm acquires Modular

Chris Lattner (creator of LLVM and Swift) announced Qualcomm's acquisition of Modular; Mojo's open-source plans continue. Non-CUDA inference stacks gain a heavyweight backer.

12. Funding and talent

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

13. Safety and governance

Anthropic is involved in export-control litigation, vulnerability-testing controversy, and distillation allegations.
The proposed US Chip Security Act would require location tracking on high-end AI chips; the community raises security and privacy concerns.
Virginia data centers face resident noise complaints (Carmack joined the discussion) — the social cost of AI infrastructure is surfacing.

14. Products and papers worth a look

OpenThoughts-Agent: 100+ ablations, a 100K-sample fine-tuning set for Qwen3-32B, averaging 44.8% across 7 agent benchmarks. Instruction quality, trajectory length, and data source all matter.
Kimi API on AWS Marketplace: convenience for enterprise procurement more than new capability.
Seedance for anime: a Japanese animator's Reddit-shared workflow uses 3D scenes for motion composition plus AI rendering, keeping long-video consistency better than pure text-to-video.
ChatGPT photo restoration exposes "identity drift": age-and-restore tests produced plausible but non-identical faces — inpainting/generation, not faithful restoration.

> 📌 Information above comes from the easy-learn-ai daily update for 2026-06-25, compiled and tracked by the community. Original source links are available in easy-learn-ai's data/daily/md/2026-06-25.md.