Motif Technologies Releases Motif-3 Beta, Its Own 314B MoE

On July 20, Motif Technologies published a beta checkpoint of its large-scale Mixture-of-Experts (MoE) language model, Motif-3, on Hugging Face. The company emphasizes that this is a fully in-house, purpose-built architecture rather than a re-parameterization of an existing open-source architecture.

The main specs are as follows:

Total parameters: ~314B; effective parameters per token: ~13B

Context length: 262,144 (256K)

routed experts (top-8) plus 1 shared expert

Proprietary components such as Grouped Differential Latent Attention (GDLA) and Grouped PolyNorm

Its current Artificial Analysis Intelligence Index (AAII) score of 44 is modest, but the final checkpoint is due to be released soon. Use is limited to personal, educational, and non-commercial research; commercial use requires written permission from the company.

From Hugging Face: A Wave of Speech AI for Low-Resource Languages

Two speech-related models are drawing attention on Hugging Face's Daily Papers.

**GigaAM Multilingual** is an MIT-licensed speech foundation model for low-resource languages. It offers two Conformer encoders, 220M and 600M, pre-trained in a HuBERT-style setup on 2 million hours of data across more than 70 languages. It achieves 32–86% relative WER reductions over Whisper-large-v3 on Central Asian languages such as Kazakh, Kyrgyz, and Uzbek, claiming state-of-the-art accuracy among open-source models. Reportedly, it can also adapt to a new language with just a few hours of labeled data.

**GigaChat Audio 10B (A1.8B)** is a speech-native LLM. It combines a GigaAM encoder with a 10B MoE decoder with 1.8B effective parameters, supporting ASR, translation, spoken QA, emotion recognition, and temporal grounding for recordings up to 2 hours long. By embedding periodic temporal anchors called 'inter-timings' into the speech stream, it reportedly achieves 48.3 mIoU on temporal localization, far ahead of other open-source models (which score nearly 0.0).

Speech AI still has far more room for improvement than text-centric LLMs, and steady progress is visible on both the multilingualization and long-form processing fronts.

Gritt, Aiming to Automate Construction Sites, Raises $34 Million

According to TechCrunch, construction robotics startup Gritt has emerged from stealth with $34 million in funding. Starting with the construction of solar power plants, it aims to automate the most difficult tasks on construction sites. The construction industry faces severe labor shortages, making it a sector that has drawn growing attention in recent years as a domain for robotics applications.

US Army Faces AI Token Shortage, Moves to Restrict Usage

According to WIRED, members of the US Army received an email urging them to limit their AI usage because AI tokens are rapidly running out. Cost control is a recurring challenge for organization-wide AI use, and this is a case that tests how to balance per-token pricing against usage volume. That even an organization as large as the military hit its cap faster than anticipated suggests generative AI is spreading on the ground at a greater pace than expected.

Anthropic Sued for Copyright Infringement — Legal Battles over LLM Training Continue

As discussed in Reddit's LocalLLaMA community, Anthropic has been sued for using copyrighted books to train its LLMs. Details of the complaint are limited in the data collected this round, but the legal battles over AI companies' use of training data continue, involving OpenAI and others — a development that could shape data-sourcing strategies across the industry and deserves close attention.

DRAM Shortage Could Last Another Decade — Hardware-Side Constraints Come into Focus

Citing statements by ADATA's chairman, reports warn that the DRAM shortage could persist for another 10 years (via TechPowerUp). The structural issue is that production capacity cannot keep up with the surging demand for AI inference and training. With models growing larger and inference infrastructure expanding, the argument that memory-side bottlenecks could constrain the AI industry's growth reflects a reality that software players alone cannot sustain.

---