Easy AI Daily Digest | December 10, 2025: Devstral 2, Agentic AI Foundation, and More

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily digest for December 10, 2025 covers major model releases and industry developments. Mistral launched Devstral 2, a 123-billion-parameter coding model with a 256K-token context window scoring 72.2% on SWE-bench Verified, plus a lighter Devstral Small 2 at 24B parameters. Zhipu released GLM 4.6V-Flash, a 10B coding model optimized for consumer GPUs. Anthropic donated the Model Context Protocol (MCP) to the newly formed Agentic AI Foundation under the Linux Foundation, with OpenAI's Agents.md and Block's Goose as founding projects. Other news includes Grok Imagine 5 image model launch, tool-calling bugs in Qwen3-VL 30B on llama.cpp, a leaked Deepseek jailbreak method, NVIDIA A100 FLOPS specification clarifications, a study finding only 10.5% of Claude's code suggestions meet security standards, and Perplexity Pro's discovered 600-query daily limit.

Easy AI Daily | December 10, 2025 AI Industry Roundup

Model Releases & Updates

Mistral Releases Devstral 2 Coding Model

Mistral's Devstral 2 is a dense 123B-parameter coding model with a 256K-token context window, scoring 72.2% on SWE-bench Verified. API costs dropped 90%, weights are open-source, and performance matches Sonnet 4.3. The lightweight Devstral Small 2 (24B parameters) scores 68% on SWE-bench.

Links: Mistral news | Hacker News discussion | Reddit thread

GLM 4.6V-Flash Coding Model Released

A 10B-parameter model optimized for coding scenarios, running at up to 70 tokens/sec on an RTX 2060 (Q4 quantization). Users report it outperforms peer models in its size class.

Links: Hugging Face model | LM Studio Discord

Chronos-1.5B Quantum-Classical Hybrid Model

Trained on IBM Heron r2 quantum hardware with an embedded 2-qubit core, combining real quantum training (not simulation) with the VibeThinker-1.5B model.

Links: Hugging Face model | IBM Quantum

Grok Imagine 5 Image Model Launches

A new image model positioned against Nano Banana Pro; reviews are mixed, though earlier Grok models were well received for generation quality and UI.

Links: LMArena Discord

Qwen3-VL 30B A3B-Instruct Tool-Calling Issues

Users report empty assistant responses when using Qwen3-VL with llama.cpp, while the non-multimodal variant works fine.

Links: Discord discussion

Agentic AI & Protocols

Agentic AI Foundation Formed Under the Linux Foundation

Anthropic's Model Context Protocol (MCP), OpenAI's Agents.md, and Block's Goose are founding projects, aiming to standardize agentic AI development.

Links: Foundation site | Block announcement

Anthropic Donates MCP to the Linux Foundation

MCP now falls under the Agentic AI Foundation's governance, with existing governance structure and workflows unchanged.

Links: Anthropic news | Linux Foundation press release

AI in Software Development

Cursor Agents bugs persist: Users report the feature failing and requiring manual file creation; some consider switching to Antigravity. (forum)
Cursor adds sub-agents, docs missing: A new sub-agent module appeared under .cursor/agents, but orchestration documentation hasn't been published yet. (Discord)
GitHub Copilot reviews polarized: Some users value it; others criticize functional defects and redundant reasoning. (Discord)

AI Ethics & Safety

Deepseek jailbreak leaked: A method using cybersecurity-themed prompts to bypass Deepseek's content moderation circulated in a community Discord. (InjectPrompt)
Desktop Commander flagged as security risk: Users warn of malicious code injection and privacy leakage risks, suspecting a scam tool. (screenshot)

Infrastructure & Hardware

A100 FLOPS specs clarified: The listed 156 TFLOPS is TF32 matrix multiply-accumulate (19-bit) performance; 312 TFLOPS is FP16 matrix multiply-accumulate; element-wise operations are comparatively slower. (NVIDIA docs)
CUDA beginner tutorial revamp: Official beginner docs are being revised, with a live walkthrough planned after the maintainer returns from parental leave. (CUDA docs)

Community Events

AI Agents 0-to-1 Workshop on Dec 13: Teaches building AI agents with Langchain and Streamlit around real client projects; registration open. (Luma)
GitHub Social Club meetup in NYC: Coffee, swag, and face time with the GitHub Copilot team in SoHo. (GitHub events)

Research & Papers

Claude code assistance raises security concerns: A study found only 10.5% of Claude's code suggestions meet security standards, with 61% functional; it also evaluated Claude Sonnet 4, Gemini 2.5 Pro, and Kimi K2. (arXiv)
SmolVLA paper corrects SO100/SO101 dataset mix-up: The SO100 dataset contains pick-place, stacking, and sorting tasks, while SO101 contains only one task type. (arXiv | SO101 dataset)

Tools & Services

Perplexity AI Pro capped at ~600 queries/day: Despite "unlimited" marketing, users found a daily cap of about 600 queries, with the limit language removed from the site. (perplexity.ai)
Optuna arrives on Hugging Face: Hyperparameter optimization support added to help optimize training scripts. (PR)

---

📌 Source: Easy AI Daily