Easy AI Daily Digest | December 30, 2025

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for December 30, 2025 covers major AI industry developments: Tencent released WeDLM 8B Instruct, a diffusion language model 3-6x faster than vLLM-optimized Qwen3-8B under Apache 2.0; fal open-sourced FLUX.2 Turbo, ranked first among open image models on Artificial Analysis; MiniMax-M2.1 topped Code Arena's open WebDev rankings; and vLLM launched its official community site vllm.ai. Performance findings include AMD MI300X FP8 underperforming bf16 on MiniMax-M2.1, while Baseten reported 20% faster GLM-4.7 inference. Research highlights: Google showed Transformers learn implicit multi-hop reasoning, URM outperforms static-depth models on ARC-AGI, TTT-E2E extends 3B model context from 8K to 128K, and AgentReuse cuts agent latency by 93%. Industry news includes Meta's ~$4B acquisition of Manus AI and xAI hiring for RL post-training and safety roles.

Models & Frameworks

vLLM launches official community site vllm.ai: Includes an interactive installation selector, event calendar, centralized documentation hub, and an office-hours playlist.
Tencent releases WeDLM 8B Instruct: A diffusion language model on Hugging Face, 3-6x faster than vLLM-optimized Qwen3-8B, Apache 2.0 licensed, with strong benchmark results. Reddit discussion
fal open-sources FLUX.2 Turbo: Based on DMD2 distillation, claimed #1 among open-source image models on Artificial Analysis; community Hugging Face Spaces demos appeared quickly. Release tweet
MiniMax-M2.1 leads open agentic coding: Ranked #1 open WebDev on Code Arena; Chutes testing showed 82.83% tool-call accuracy, with iterations toward M2.2/M2.5.

Inference & Performance

AMD MI300X FP8 underperforms bf16 on MiniMax-M2.1: vLLM bf16 reached 55.7 TPS (FP8: 42 TPS); sglang bf16 71 TPS (FP8: 55 TPS).
Weaviate new release: Object TTL, Java v6 client GA, Flat Index RQ quantization, zstd backups, multimodal document embedding.
Baseten reports GLM-4.7 inference 20% faster (tok/s and TTFT); GLM-4.7 is now their internal default coding model.

Open Models & Datasets

GLM-4.7 cited by AlphaXiv as #1 on Artificial Analysis for open coding models.
pokeart dataset: 1,224 Pokémon splash art and battle sprites (Gen1–Gen9) with captions from Gemini 3 Pro and Qwen3, on Hugging Face.
Korean 32B VLM released with architecture tweaks (muP and sandwich norm removed, 0.006 init), strong English/Korean benchmarks; technical report pending.

AI Agents & Workflows

Spotify's production coding agent lessons: specify verifiable end states, include code examples, minimize tools (verify/git/bash), document workflows in AGENTS.md.
Dual-audience documentation pattern for AI agents discussed; LlamaIndex offers templates.
Amazing Z-Image Workflow v3.0 released: Style Selector (15 styles), Sampler Switch, Z-Image Enhancer, GGUF/safetensors support (GitHub).
OpenEnv (Meta × Hugging Face) standardizes agent environments for frameworks like TRL/TorchForge with MCP tool integration.

Research Highlights

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

Industry & Acquisitions

Meta acquires Manus AI for ~$4B: Manus reached $100M ARR in 9 months; Alexandr Wang announced the deal, citing SOTA performance on Remote Labor Index.
xAI hiring for RL post-training, alignment/behavior, and catastrophic risk reduction roles.
Perplexity Pro limits reported: users note caps on advanced models (~1-2/hour), while Max tier is advertised as unlimited; some subscription issues reported.

Community Discussions

AI-generated image tells (trays as feet, warped glasses) discussed on Reddit; commenters joked that correct toe counts now look suspicious.
OpenAI's "Killswitch Engineer" job posting meme debated as marketing, with skepticism about the $300-500K salary for a "pull the plug" role.
BASI Jailbreaking members found XSS vulnerabilities in a vibe-coded app's LLM-generated JavaScript, discussing input validation.
Unsloth AI members discussed multi-head attention subspaces and training data as probabilistic compression of LLM knowledge.
Perplexity users explored open-source alternatives like Perplexica.