Easy AI Daily Digest — December 9, 2025

A curated roundup of AI industry news originally published on zhichai.net by the Easy AI teaching project.

Key points

Model updates & multimodal progress

• Zhipu releases GLM-4.6V and GLM-4.6V-Flash: GLM-4.6V is a 106B MoE model for cloud/high-performance clusters; GLM-4.6V-Flash is a 9B dense model for local, low-latency use. Both support 128k context and native multimodal function calling. Flash API is free, weights open on Hugging Face (GLM-4.6V-Flash, GLM-4.6V).
• Jina AI releases Jina-VLM-2B: A 2B multilingual VLM focused on charts, documents, and scene text. Averages 72.3 across 8 VQA benchmarks with SOTA on MMMB (78.8) and multilingual MMBench (74.3) (announcement).
• Qwen3-4B impresses locally: ~70 tokens/sec on an RTX 2060 with strong coding, though users note overthinking behavior; recommended to pair with GLM-4.6V-Flash.
• DeepSeek V3.2 improves reasoning: Supports interleaved reasoning, strong RoO code performance, priced at $0.28/M input and $0.45/M output tokens, more stable than Kimi.

Training & optimization tools

• Hugging Face Claude Code skills automate LLM training: Specify tasks in natural language (e.g., fine-tuning Qwen3-0.6B); handles data validation, GPU selection, and HF Jobs submission; small runs cost ~$0.30 (blog).
• Unsloth AI updates: Celebrates 10k Reddit members, fixes slow HF downloads, and releases Mistral Large 3 GGUF (675B) for local inference (GGUF).
• DSPy TOON Adapter: Reduces token consumption but handles nested schemas less well than BAMLAdapter.
• LangChain Deep Agents evaluation: Framework for long-running agent evaluation; Terminal Bench 2.0 mean of 42.65%, with context compression triggers (tweet).

Research & evaluation

• Google's Miras framework (NeurIPS): Treats Transformers/RNNs as associative memory systems; outperforms Transformer, Mamba2, and DeltaNet on LM/reasoning/long-context tasks with 20% gains in long-text retrieval (overview).
• AxiomProver solves 9/12 Putnam problems: Lean-based system with formalized, verifiable pipelines, outperforming last year's top human score (release).
• NeurIPS mechanistic interpretability workshop: Chris Olah argues for scalable interpretability tools over single-model neuron analysis.
• MEMTRACK benchmark: Evaluates agent long-term memory in Slack/Linear/git scenarios; GPT-5 scored 60%, revealing room for improvement.

Agents & workflows

• Dexter 2.0: Open-source financial research agent with planning and self-verification, built on LangChain (demo).
• AI21 Maestro: Agent orchestration with multi-step planning, built-in validation, proprietary RAG, and execution graphs (release).
• Cursor agent issues: Reports of failure to create files and infinite loops; temporary workaround via approval button.

Infrastructure & hardware

• OpenAI Stargate strains DRAM supply: Consumes up to 40% of global DRAM output (~900k wafers/month) via Samsung and SK Hynix deals, driving DDR5 price increases (Tom's Hardware).
• AMD 7900 XTX as budget AI GPU: Strong price/performance and good llama.cpp support for local LLMs (discussion).
• Blackwell WGMMA incompatibility: Blackwell GPUs lack WGMMA (sm90a only), requiring kernel rewrites; CUDA 13.1 introduces CUDA Tile (blog).
• RadixArk spins off from SGLang: Focuses on AI scheduling, compilers, serving, and training pipelines (announcement).

Community & ecosystem

• Local LLM cluster build: 8x RTX 3090 (192GB VRAM) + 64-core EPYC Milan + 250GB RAM for ~$8k; runs GLM-4.5 Air Q6_K at 49 tokens/sec via llama.cpp (details).
• Vector DB selection guide: HNSW for <10M vectors, Turbopuffer for large datasets, pgvector for small/local setups, Chroma for lightweight use (guide).
• Key community contributors recognized: Unsloth (fast fine-tuning), mradermacher (automated quantization), Bartowski (curated quants), TheBloke (base models).
• picomon: Open-source AMD GPU monitoring tool from an HF user (repo).

Industry & market

• Meta acquires Limitless (formerly Rewind): Pendant users get 1 year of support; non-Pendant features will sunset.
• ARC Prize 2025 results: NVARC tops the board at 25.03%; the TRM paper wins $50k; the $600k Grand Prize remains unclaimed (results).
• Manus.im billing problems: Users report lost credits and auto-renewals, suspected system bug.
• Sora 2 regional restrictions: Available in only 7 countries; VPN use violates ToS and may lead to account bans (supported countries).

Fun & memes

• ChatGPT "napping on the job" screenshots go viral; an AI bug-fixing loop meme resonates with developers; a David Duchovny AI image with 20 verified details and a Nano Banana Pro fashion contact-sheet workflow gain community praise.

*Source: Easy AI teaching project (zhichai.net)*