Easy AI Daily News Digest | December 30, 2025

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for December 30, 2025 covers major AI industry updates: Tencent released the WeDLM 8B Instruct diffusion language model (Apache 2.0, 3-6x faster than vLLM-optimized Qwen3-8B); fal open-sourced FLUX.2 Turbo, a DMD2-distilled image model ranked first among open models on Artificial Analysis; MiniMax-M2.1 topped open-source WebDev rankings on Code Arena; and vLLM launched its official community site vllm.ai. Benchmarks showed AMD MI300X bf16 outperforming FP8 for MiniMax-M2.1. Research highlights include Google findings that Transformers store global structure enabling multi-hop reasoning, URM's 53.8% pass@1 on ARC-AGI 1, TTT-E2E extending 3B-model context from 8K to 128K, and AgentReuse cutting agent latency by 93%. Industry news: Meta acquired Manus AI for roughly $4 billion, xAI opened RL post-training and safety roles, and users reported Perplexity Pro usage limits.

📅 AI Industry Digest — December 30, 2025

Models & Frameworks

vLLM launches official community site vllm.ai — includes an interactive installation selector, events calendar, and centralized documentation hub, with acknowledged documentation gaps addressed via search and office-hours playlists.
vLLM release tweet | Docs tweet
Tencent releases WeDLM 8B Instruct — an open-source diffusion language model on Hugging Face, claimed 3-6x faster than vLLM-optimized Qwen3-8B, Apache 2.0 licensed, strong benchmark results.
Hugging Face | Reddit discussion
fal open-sources FLUX.2 Turbo — DMD2-distilled image model, claimed #1 open model on Artificial Analysis; community Hugging Face Spaces demos appeared quickly.
fal tweet | Demo
MiniMax-M2.1 — ranked #1 open-source WebDev on Code Arena; Chutes tests showed 82.83% tool-calling accuracy; iteration toward M2.2/M2.5 planned.
Code Arena | Chutes tests

Inference & Performance

AMD MI300X: FP8 underperforms bf16 — Multiple reports show bf16 beating FP8 for MiniMax-M2.1: vLLM bf16 55.7 TPS vs FP8 42 TPS; sglang bf16 71 TPS vs FP8 55 TPS.
QuixiAI test 1 | Test 2
Weaviate new release — Object TTL, Java v6 client GA, Flat Index RQ quantization, zstd backups, multimodal document embeddings. Tweet
Baseten optimizes GLM-4.7 — 20% faster inference (tok/s and TTFT); GLM-4.7 became Baseten's internal default coding model.
Tweet 1 | Tweet 2

Open Models & Datasets

GLM-4.7 becomes open-source coding default — via Interleaved/Preserved/Turn-level Thinking; AlphaXiv reports #1 on Artificial Analysis. AlphaXiv
pokeart dataset — 1,224 Pokémon (Gen1–Gen9) splash art and battle sprites with 6 Gemini 3 Pro captions and 1 Qwen3 caption, for benchmarking. Hugging Face
Korean 32B VLM released — architecture tweaks (removed muP and sandwich norm), 0.006 initialization, strong English/Korean benchmarks; technical report pending. Tweet

AI Agents & Workflows

Spotify's coding agent production lessons — specify verifiable end states, include code examples, minimize tools (verify/git/bash), document workflows in AGENTS.md.
Tweet 1 | Tweet 2
Dual-audience documentation — docs for both humans and agents (AGENTS.md/CLAUDE.md conventions); LlamaIndex offers templates. Tweet
Amazing Z-Image Workflow v3.0 — Style Selector (15 styles), Sampler Switch, Landscape Switch, Z-Image Enhancer; supports GGUF/SAFETENSORS.
Reddit | GitHub
OpenEnv (Meta × Hugging Face) — standardizes agent environments, supports TRL/TorchForge, integrates MCP tools to unify training and deployment.
Tweet 1 | Tweet 2

Research Highlights

Transformers store global structure — Google research shows implicit multi-hop reasoning with 100% accuracy on 50k-node graphs, challenging knowledge-editing assumptions.
dair.ai
URM (Universal Recurrent Model) — recursive inference beats static depth; 53.8% pass@1 on ARC-AGI 1 via recurrent inductive bias and strong nonlinearity. Tweet
TTT-E2E — test-time training compresses context, extending a 3B model's context from 8K to 128K, 2.7x faster than full attention with better performance. Tweet
AgentReuse — caches and parameterizes agent plans; 93% reuse rate across 2,664 requests, 93% latency reduction, low memory overhead. Tweet

Industry & Acquisitions

Meta acquires Manus AI for ~$4 billion — Manus reached $100M ARR in 9 months; Alexandr Wang announced the acquisition and Manus joining Meta Superintelligence Labs, citing SOTA on the Remote Labor Index.
Tweet 1 | Tweet 2 | scaling01
xAI hiring — RL post-training, alignment/behavior, and catastrophic risk reduction roles. Tweet
Perplexity Pro limits — users report premium-model caps (1-2/hour) despite the Max tier advertising unlimited use; some students saw subscriptions terminated after 7 months of a 12-month voucher.
Reddit 1 | Reddit 2

Reddit & Community Highlights

Spotting AI-generated images — anomalies like trays with feet instead of handles and distorted glasses are common tells; users joked that correct toe counts are now suspicious. (GeminiAI | StableDiffusion)
OpenAI "Killswitch Engineer" meme — users debated the viral job-title meme questioning the legitimacy of a $300k–$500k "unplug the plug" role. (OpenAI | ChatGPT)
Discord community — BASI Jailbreaking found XSS vulnerabilities in vibe-coded apps when LLMs generate payloads; Unsloth AI discussed multi-head attention subspaces and the importance of covering edge cases in training data; Perplexity users explored open-source alternatives like Perplexica.

---

📌 Source: Easy AI Daily 🤖 Compiled by: AI assistant