Easy AI Daily Digest | December 9, 2025: Model Releases, Agent Tools, and Industry News

📅 AI Industry Digest — December 9, 2025

*English translation of the Easy AI Daily digest from zhichai.net.*

Model Updates & Multimodal Progress

Zhipu releases GLM-4.6V and GLM-4.6V-Flash multimodal models

Zhipu launched GLM-4.6V (106B MoE, for cloud/high-performance clusters) and GLM-4.6V-Flash (9B dense, for local low-latency use). Both support 128k context and native multimodal function calling. The Flash API is free, and weights are available on Hugging Face.

Links: GLM-4.6V-Flash on Hugging Face | GLM-4.6V on Hugging Face

Jina AI releases Jina-VLM-2B

A 2B-parameter multilingual vision-language model focused on charts, documents, and scene text. It averages 72.3 across 8 VQA benchmarks, with SOTA on MMMB (78.8) and multilingual MMBench (74.3).

Link: Release tweet

Qwen3-4B performs well locally

Reaches 70 tokens/sec on an RTX 2060 with strong coding ability, though users note an "overthinking" issue and recommend pairing it with GLM-4.6V-Flash.

DeepSeek V3.2 improves reasoning

Supports interleaved reasoning with strong RoO code performance. Pricing: $0.28/M input tokens, $0.45/M output tokens. Reported as more stable than Kimi.

Training & Optimization Tools

Hugging Face Claude Code skill automates LLM training

Specify training tasks in natural language (e.g., fine-tuning Qwen3-0.6B); the skill handles data validation, GPU selection, and HF Jobs submission, with small runs costing roughly $0.30.

Links: HF skills training blog | LiorOnAI tweet

Unsloth AI updates

Celebrated 10k Reddit members, fixed slow HF downloads, and released Mistral Large 3 GGUF models for local inference.

Links: Unsloth Reddit | Mistral Large 3 GGUF

DSPy TOON Adapter reduces token usage

The community-released TOON Adapter cuts token consumption but handles nested schemas less well than BAMLAdapter; GEPA optimization improved MMLU-Pro results.

Link: TOON Adapter code

LangChain Deep Agents evaluation tooling

An evaluation framework for long-running agents: Terminal Bench 2.0 mean of 42.65%, plus context-compression triggers.

Link: LangChain tweet

Research & Evaluation

Google's Miras post-Transformer framework

A NeurIPS paper framing Transformers/RNNs as associative memory systems. Miras outperforms Transformer, Mamba2, and DeltaNet on LM, reasoning, and long-context tasks, with 20% improvement in long-text retrieval.

Link: Introduction tweet

AxiomProver solves 9/12 Putnam problems

The Lean-based system solved 9 problems within hours of the exam, emphasizing verifiability and formal pipelines, surpassing last year's top score.

Link: Release tweet

NeurIPS mechanistic interpretability workshop

Chris Olah reflected on interpretability, arguing the field needs scalable tools rather than neuron-level analysis of single models; the community echoed calls for more general methods.

Link: NeelNanda5 tweet

MEMTRACK benchmark for agent long-context memory

Evaluates agent memory in Slack/Linear/git scenarios; GPT-5 scored 60%, showing room for improvement in real tool environments.

Link: Release tweet

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; approval button is a temporary workaround. Discussion

Infrastructure & Hardware

• OpenAI Stargate and DRAM shortage: Stargate may consume up to 40% of global DRAM output (~900k wafers/month, deals with Samsung and SK hynix), pushing up DDR5 prices. Report
• AMD 7900 XTX as a budget AI GPU: strong price/performance and good llama.cpp support for local LLMs. Discussion
• Blackwell WGMMA incompatibility: WGMMA is sm90a-only, causing compile errors; CUDA 13.1 introduces CUDA Tile to simplify programming. CUDA 13.1 blog
• RadixArk spun out of SGLang, focusing on AI scheduling, compilers, serving, and training pipelines. Tweet

Community & Ecosystem

• Local LLM cluster: a user built an 8x RTX 3090 (192GB VRAM) + 64-core EPYC Milan + 250GB RAM rig (~$8k), running 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 databases, pgvector for small/local setups, Chroma for lightweight use. Blog
• Community contributors: Unsloth (fast fine-tuning), mradermacher (automated quantization), Bartowski (curated quants), and TheBloke (base models) recognized as core contributors. Discussion
• picomon: open-source AMD GPU monitoring tool from an HF user, more reliable than nvtop at some cost in precision. Code

Industry & Market

• Meta acquires Limitless (formerly Rewind); Pendant users get 1 year of support, non-Pendant features will sunset. Tweet
• ARC Prize 2025 results: NVARC took Top Score at 25.03%; the TRM paper won $50k; the $600k Grand Prize went unclaimed; winning methods will be open-sourced. Results
• Manus.im billing issues: users report lost credits and subscriptions auto-renewing into 2026, with slow support responses; suspected system bug. Discussion
• Sora 2 regional restrictions: available in only 7 countries; VPN use violates the ToS and may lead to account bans. Supported countries

Fun & Memes

• ChatGPT "napping on the job" screenshots went viral. Post
• A meme about AI endlessly claiming a bug is fixed resonated with developers. Video
• A hyper-detailed David Duchovny AI image with a 20-item checklist verifying each detail. Tweet
• A Nano Banana Pro contact-sheet prompt workflow generating 6-frame fashion editorials with camera and styling constraints. Details

---

*Source: Easy AI educational project.*