IBM Granite 4.1 30B Arrives as Open Source, DeepMind Co-founder Raises $1.1B, and Small LLMs See Dramatic Gains from Self-Refinement

IBM Granite 4.1 30B — A Multilingual, Long-Context Open Model

IBM has released **Granite 4.1-30B** on Hugging Face. It is a 30B-parameter dense Transformer that supports long contexts of up to 131K tokens.

Its standout feature is **multilingual support**, covering 12 languages including English, Japanese, German, Spanish, French, and Chinese. After supervised fine-tuning (SFT) and reinforcement-learning-based alignment, the model reportedly delivers major improvements in tool calling, instruction following, and chat performance.

On benchmarks, it scores 80.16 on MMLU 5-shot and 88.41 on HumanEval pass@1. It also achieves 73.68 on the BFCL v3 tool-calling tasks, showing strong real-world capability. The model was trained on an NVIDIA GB200 NVL72 cluster (CoreWeave) and is available under the Apache 2.0 license.

DeepMind Co-founder David Silver Raises $1.1B — AI Startup "Ineffable"

**David Silver**, one of the co-founders of DeepMind, has launched a new AI startup called **Ineffable** and raised **$1.1 billion (approximately ¥160 billion)** in seed funding. NVIDIA and Google are among the investors.

Silver led groundbreaking reinforcement learning projects such as AlphaGo and AlphaZero. What specific AI technology Ineffable will develop has not been disclosed at this point, but a seed round of this scale is highly unusual and is drawing significant attention across the AI industry.

Small LLMs Dramatically Improved with a "Bidirectional Refinement Loop"

A report on Reddit's LocalLLaMA community is gaining attention, claiming that adding a mechanism called **"Second Thoughts"** to a small 1.7B-parameter model achieved substantial accuracy gains on code generation tasks.

The mechanism is simple: a small transformer is added to read the output near the end of generation, and its output is fed back to near the beginning of the input, forming a loop. This reportedly enables the model to self-correct grammatical and syntactic errors. The idea is inspired by insights from neuroscience, and additional validation on a 9B model is said to be underway.

Running Mistral Medium 3.5 128B Locally on 3× RTX 3090s

Also on LocalLLaMA, results were reported from running **Mistral Medium 3.5-128B** locally with Q3_K_M quantization on 3× RTX 3090s (72GB VRAM).

The report verifies actual inference speed and output quality across Python code generation, SVG, and HTML tasks. The finding that even a 128B-class model can run at practical speeds on consumer GPUs is an interesting data point illustrating how capable local LLMs have become.

Themis — A Multilingual, Multi-Criteria Code Reward Model

**Themis**, which attracted attention on Hugging Face Daily Papers, is a research effort proposing a new approach to reward models for code generation.

While conventional code RMs have been biased toward the single criterion of "does it run," Themis enables evaluation across **five criteria** (readability, efficiency, security, and more) and **eight programming languages**. The team built the Themis-CodePreference dataset consisting of over 350k pairs and released a family of models ranging from 600M to 32B parameters, demonstrating that training on multiple criteria is essential for reliable code evaluation.

Big Tech Cuts 80,000 Jobs and Cites AI, but Experts Point to "Overstaffing"

Past **layoffs of roughly 80,000 employees** at big tech companies have been attributed to AI adoption, but experts argue that these companies were actually 25–75% overstaffed. While AI-driven replacement is part of the picture, analysts note that correcting structural overhiring is also a major part of the context.

---