NVIDIA Releases NVFP4-Quantized GLM-5.2 — Better Inference Efficiency with Accuracy Largely Intact

On June 25, NVIDIA released GLM-5.2-NVFP4 on Hugging Face, an NVFP4-quantized version of GLM-5.2, the large language model from ZAI (formerly Zhipu AI). GLM-5.2 adopts a Mixture-of-Experts (MoE) architecture and is a reasoning- and coding-focused model boasting 753B total parameters (40B active) and a maximum context length of 1M tokens.

The accuracy loss from quantization is extremely small, with the model maintaining results on par with the FP8 baseline on benchmarks such as GPQA Diamond (89.52 → 89.39) and SciCode (49.85 → 49.04). On IFBench, AA-LCR (long-context recall), and τ²-Bench Telecom (agentic tool use), the quantized version even recorded results exceeding the baseline.

It is intended for inference with SGLang and vLLM, and has been tested on NVIDIA B200/B300 environments. It is offered under the MIT license for both commercial and non-commercial use, keeping the barrier to production adoption low.

NVIDIA Nemotron-3-Super-120B — Hybrid Mamba Architecture Fully Processes 500K Tokens

Another NVIDIA model, Nemotron-3-Super-120B-A12B, is attracting attention in the LocalLLaMA community. It employs a unique architecture (nemotron_h) combining hybrid Mamba2 + periodic attention + MoE, designed so that the Mamba layers hold a fixed-size recurrent state, preventing KV cache bloat.

According to hands-on reports from the community, on a setup with 4× RTX 3090 GPUs (96GB total VRAM, about 71GB used), the model achieved perfect recall at all depths on needle-in-a-haystack tests up to 504,482 tokens. Decode speed held at 23 t/s even at 500K tokens, comparable to the speed that similarly sized full-attention MoE models record at the 30K-token mark.

On the other hand, a "recency bias" has also been observed, where instructions embedded in the first half of a document get overridden by contradictory instructions in the latter half. The shared takeaway: in long documents, hard rules should be placed at the end or in the system prompt.

Anthropic Moves Toward Agreement with US Government on Easing AI Model Restrictions

According to a Bloomberg report on June 26, Anthropic is moving toward an agreement with the US government to ease restrictions on AI models. The details need to be confirmed in the article itself, but this is seen as a significant development suggesting that the relationship between AI companies and regulators is changing.

In recent years, over AI model export controls and security requirements, major companies including OpenAI and Anthropic have been in active negotiations with governments. We await future official announcements on what specific conditions this reported deal includes.

Japan's Physical AI Investment Reaches ¥10.5 Trillion — "From Proof of Concept to Deployment"

According to MONOist, combined public and private investment in physical AI (AI systems operating in real physical environments) in Japan has reached ¥10.5 trillion. The article highlights how the phased transition from proof of concept to deployment is now underway in the field.

Physical AI covers a broad range of areas, including robotics, autonomous driving, and manufacturing automation. For the labor shortage issues facing Japanese society, entering the deployment phase means an acceleration of practical adoption.

Local JSON Extraction from Receipts with Qwen3.6-35B-A3B — A Demonstrated Alternative to Google Vision

The LocalLLaMA community has reported a proof-of-concept experiment running Qwen3.6-35B-A3B on an RTX 3060 (12GB) to extract the store name, date, subtotal, tax amount, and total from Japanese receipt images in JSON format.

In tests on about 30 Japanese receipts, the fields of interest were extracted consistently and accurately. Processing time was about 31.75 seconds per receipt, with peak VRAM usage of around 11GB. This demonstrated that it can be a practical alternative to the Google Vision API from both privacy and cost perspectives.

While it falls short of cloud APIs in terms of latency, the advantage of not sending data externally is significant, and the approach is applicable to extracting standardized documents such as invoices and forms as well.

---