AI Technology Observations | 29 October 2026, 07:00

This observation prioritises AI technology changes formally disclosed or reliably verified from 07:35 on 28 October to 07:35 on 29 October 2026 in Australia/Sydney time. It also includes one actual release disclosed immediately before the window and not previously recorded by this series. Three items are included.

NVIDIA releases Nemotron 3.5 Transcribe as an open-weight speech model

NVIDIA released Nemotron 3.5 Transcribe on 28 October together with BF16, FP8, NVFP4 and GGUF checkpoints, a DFlash2 draft model and NeMo Speech fine-tuning tools. The model has about 34 billion total parameters and activates about 4.3 billion per inference pass. It accepts recorded audio and text instructions, covers 40 languages, supports recordings up to one hour and a 65,536-token context, and can perform verbatim or written-form transcription, translation, speaker attribution, word-level timestamps and language identification. The weights use the OpenMDW-1.1 licence. The model card confirms downloadable commercial and non-commercial use, subject to that licence and applicable third-party component terms.

The released system is an open-weight offline recording model and does not support live audio streams. English and multilingual ASR, speaker-transcription and single-GPU throughput figures are NVIDIA-reported. TechCrunch confirmed that the model, weights and Granary v2 speech dataset were released but did not independently reproduce performance. The model card also states that the optimised serving path in NVIDIA NIM has not yet been upstreamed to the open-source release and that the release-day open vLLM integration is expected to be roughly twice as slow, so hosted-service throughput should not be treated as a local-deployment result.

Liquid AI releases the compact multimodal LFM3 family

Liquid AI released the first six LFM3 models, each around two billion parameters, on 27 October, covering text, vision, audio, omni input and two structured-decision variants. The weights are available through Hugging Face, with Playground and API access also provided. The family shares a 1.7B-parameter foundation and uses a tri-hybrid architecture combining short convolutions, sliding-window attention and global attention. The text model documents a 256K-token model context, an approximately 1 GB four-bit weight file and single-GPU fine-tuning. Access is governed by the LFM Open License v1.0 rather than an unconditional public-domain grant.

Compared with LFM2, the confirmed change is a new common backbone, multimodal branches, downloadable weights and a local fine-tuning path. Performance results and the 22-trillion-token training figure are developer-reported. VentureBeat did not independently reproduce them and found that the hosted API lists a 128,000-token context limit, below the downloadable model’s 256K specification. The official model card also says the 1.7B text model is not recommended for agentic coding or complex multi-step reasoning, so browser and tool-use benchmark results do not establish a general advantage across tasks.

The UK publishes the Data Centre Covenant

HM Treasury formally published the Data Centre Covenant on 28 October, setting conditions under which strategic data centres can receive accelerated government support. Projects that meet strategic-value and responsibility tests may receive coordinated reservation or reallocation of grid capacity, identification of capacity available before 2030, transfers of some existing connection contracts and, under tightly controlled conditions, temporary on-site generation while awaiting a grid connection. Projects must cover their energy-system costs, prepare local benefit plans, support UK jobs and hardware supply chains, and strengthen security for critical services. New facilities should also achieve a power usage effectiveness of 1.3 or better and report energy and water use annually. Reuters independently verified the Budget announcement, priority-grid and on-site-generation offer, and community obligations.

This is a published government support framework rather than approval of a particular project. The government has not yet issued the full definition of a strategic data centre or all additional eligibility details, and temporary generation is limited to exceptionally strategic projects that cannot obtain capacity in time. The draft National Policy Statement for data centres released on the same day remains under consultation. The confirmed change concerns connection and planning conditions; it does not establish new capacity, construction dates or operating scale.

Sources

NVIDIA Nemotron Research, “NVIDIA Nemotron 3.5 Transcribe: Intelligent Transcription with a Speech LLM”, 28 October 2026, https://research.nvidia.com/labs/nemotron/NVIDIA-Nemotron-3.5-Transcribe/

NVIDIA, “NVIDIA-Nemotron-3.5-Transcribe-30B-A4B Model Card”, 28 October 2026, https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B

TechCrunch, “Voice AI in vogue: Nvidia releases new open transcription model”, 28 October 2026, https://techcrunch.com/2026/10/28/voice-ai-in-vogue-nvidia-releases-new-open-transcription-model/

Liquid AI, “Introducing LFM3: Every Modality for Specialized Intelligence”, 27 October 2026, https://www.liquid.ai/blog/lfm3

Liquid AI, “LFM3-1.7B Model Card”, 27 October 2026, https://huggingface.co/LiquidAI/LFM3-1.7B

VentureBeat, “Liquid AI’s new LFM3 models put browser agents, voice commands, and image recognition onto enterprise-controlled devices”, 27 October 2026, https://venturebeat.com/technology/liquid-ais-new-lfm3-models-put-browser-agents-voice-commands-and-image-recognition-onto-enterprise-controlled-devices

HM Treasury, “The Data Centre Covenant”, 28 October 2026, https://www.gov.uk/government/publications/data-centre-covenant/the-data-centre-covenant

Reuters/Channel NewsAsia, “UK's Healey offers grid access incentives for critical data centres”, 28 October 2026, https://www.channelnewsasia.com/business/uks-healey-offers-grid-access-incentives-critical-data-centres-6500146