Easy AI Daily News | October 24, 2025: vLLM Nemotron Support, MiniMax M2, Mistral AI Studio, Karpathy's nanochat

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for October 24, 2025 covers key AI industry developments across models, platforms, research, and open source. vLLM announced support for NVIDIA's Nemotron family including the 9B Nemotron Nano 2 with hybrid Transformer-Mamba design and 6x faster thinking-token generation. MiniMax M2 entered the LMArena leaderboard, competing with Sonnet 4.5 as a low-latency agent/coding model. Mistral launched AI Studio, a production-grade agent platform with full lifecycle observability. Baseten boosted GPT-OSS 120B to 650 TPS with 0.11s TTFT. Stanford researchers proposed a 'palimpsest' method for black-box detection of model provenance with p<1e-8 statistical significance. Anthropic and collaborators introduced ImpossibleBench to test agent reward hacking. Andrej Karpathy released nanochat, an open-source end-to-end ChatGPT-like stack. Community highlights include ChatGPT aiding a 20-year diagnosis, student cheating controversies, and user complaints about Cursor Ultra billing, Perplexity referral payouts, and Manus platform reliability.

📅 AI Industry Daily — October 24, 2025

Model Updates and Releases

#### vLLM Announces Support for NVIDIA Nemotron Family vLLM now supports the NVIDIA Nemotron series, including the new 9B "Nemotron Nano 2" (hybrid Transformer-Mamba design, open weights, trained on 9T+ tokens of open data). Under vLLM, it generates "thinking" tokens 6x faster than comparable models, with long-context support and KV cache optimization. > Source: vLLM announcement

#### MiniMax M2 Lands on LMArena with Open Preview Early testing shows MiniMax M2 competing with Sonnet 4.5. It has entered the LMArena leaderboard, with usage examples available on the Yupp platform. Positioned as a low-latency, low-cost agent/coding model. > Sources: LMArena announcement | Yupp examples

#### Zhipu GLM-4.6-Air Focuses on Reliability and Infrastructure GLM-4.6-Air is still in training with reliability as a priority. Infrastructure is being expanded due to GLM Coding usage growth; users anticipate parameter efficiency gains. > Source: Zhipu update

#### Pacific-Prime Upgraded to 1.1B Parameters Pacific-Prime now reaches 1.1B parameters, delivering 10% better performance on 6GB VRAM, with claimed "zero forgetting" to preserve conversational detail. Available on HuggingFace. > Source: HuggingFace model page

#### Tahoe-x1 Single-Cell Foundation Model Released Tahoe-x1 (3B parameters) achieves SOTA on cancer-related cell biology benchmarks, unifying gene/cell/drug representations. Open-sourced on HuggingFace. > Source: Tahoe announcement

---

Platform and Tool Ecosystem

#### Mistral AI Studio: Production-Grade Agent Platform Mistral launched AI Studio, offering an agent runtime and full-lifecycle observability to help developers move from experimentation to production. > Source: Mistral announcement

#### Baseten Boosts GPT-OSS 120B Performance Baseten's GPT-OSS 120B reaches 650 TPS and 0.11s TTFT (a 44% improvement), with 99.99% uptime. Performance details and configurations published. > Source: Baseten announcement

#### InspectAI Adds Multi-Provider Model Evaluation Hugging Face InspectAI adds "inference providers" integration, enabling apples-to-apples evaluation across open model providers. > Source: InspectAI update

#### GitHub Copilot Embedding Model Improved GitHub's new Copilot embedding model improves retrieval accuracy by 37.6%, doubles throughput, and shrinks index size 8x, optimizing VS Code code search. > Source: GitHub announcement

#### Cursor Ultra Users Complain About Billing and Features Users report inaccurate budget estimates ($400 budget exhausted in a day), default PowerShell breaking Git Bash workflows, and slow support responses. > Source: Cursor community discussion

---

Research and Safety

#### Stanford Proposes "Palimpsest" Model Provenance Tracking Stanford research uses "palimpsest" metadata from training data order to black-box detect whether one model was derived from another, with statistical significance of p<1e-8 — applicable to IP protection. > Source: Research paper

#### ImpossibleBench Tests Agent Reward Hacking Anthropic and collaborators propose ImpossibleBench, using "impossible tasks" to test whether agents bypass rules (e.g., fabricating unverifiable results), improving tool-use robustness. > Source: ImpossibleBench paper

#### Sparse Memory Fine-Tuning Improves Continual Learning Jessy Lin et al. propose sparse memory fine-tuning, reducing catastrophic forgetting via dynamically activated sparsity — more efficient than LoRA under hardware bottlenecks. > Source: Research paper

#### BAPO Improves RL Post-Training Stability Fudan University proposes BAPO (dynamic PPO clipping), improving off-policy RL stability: a 32B model scores 87.1 on AIME24, and a 7B model gains 3-4 points over SFT. > Source: BAPO paper

#### Linking Transformers and Graph Neural Networks Research connects Weisfeiler-Lehman graph refinement with Transformer attention, explaining attention's structural reasoning capabilities. > Source: Research paper

---

Community and User Feedback

#### ChatGPT Helps Diagnose a 20-Year-Old Ailment After providing symptoms, test results, and medications, a user received a list of potential causes from ChatGPT and was diagnosed following up on its suggestions — sparking discussion on AI-assisted medical diagnosis. > Source: Reddit discussion

#### Template Apologies from Students Caught Cheating with ChatGPT A Reddit user shared near-identical apology emails from students caught using ChatGPT to cheat, highlighting AI's challenge to academic integrity. > Source: Reddit discussion

#### Perplexity Referral Program Controversy Users complain about missing $5 referral payouts and untracked referral leads, with accusations the platform pushes Comet Browser adoption. > Source: Perplexity Discord discussion

#### Multiple User Complaints on Manus Platform Users report network errors, fast credit burn (15,000 credits per project), outdated generated code, and unimplemented Room database support; Claude Code recommended as an alternative. > Source: Manus Discord discussion

#### LocalLlama Debates Model Reliability and Limits Users discuss GLM-4.6-Air's reliability-first strategy and Apple models being too cautious to generate random numbers. > Source: LocalLlama discussion

---

Open Source and Multimodal Projects

#### Karpathy Releases nanochat Andrej Karpathy launched nanochat, an end-to-end ChatGPT-like stack emphasizing readability and modifiability, with guidance for adding capabilities (e.g., counting letters), supporting SFT and RL optimization. > Source: nanochat announcement

#### OCR Models Gain Popularity in vLLM and Hugging Face OCR models are trending thanks to 1-click deployment (HF Inference Endpoints, vLLM). Merve published fine-tuning tutorials for Kosmos2.5 and Florence-2. > Source: vLLM OCR announcement

#### Qwen3-VL Fine-Tuned for Medieval Languages Qwen3-VL-2B/4B/8B fine-tuned on the CATmuS dataset for medieval languages/scripts, open-sourced on HuggingFace for cultural heritage applications. > Source: HuggingFace model page

#### DSPy Emerging as a LangChain Alternative Teams migrating from LangChain to DSPy cite better structured-task handling and easier model upgrades without prompt rewrites. Community also released an aider-ce fork. > Source: DSPy Discord discussion

#### LlamaIndex Supports AWS Bedrock AgentCore Memory LlamaIndex Agents integrate AWS Bedrock AgentCore Memory, providing secure storage, access control, and long/short-term memory management. > Source: LlamaIndex announcement

---

*Source: Easy AI educational project*