AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Large Language Models

Granite Team Announces Granite 4.2 Reasoning LLMs and Granite Speech 5.0 Turbo CTC Models

Granite Team released Granite 4.2 LLMs and Turbo CTC speech models featuring 12,600 RTFx speed and Apache 2.0 licensing.

AI News Feed · 2026-08-26 03:13

The Granite 4.2 post describes a family of three sizes—3B, 8B, and 30B—pre-trained from scratch on roughly 15 trillion tokens. A five-phase training strategy extends the context window to 512K tokens. The models are supervised fine-tuned on chain-of-thought, reasoning, and agentic-trajectory data, and then post-trained with a multi-stage reinforcement learning pipeline. All three models are released under the Apache 2.0 license.

Architecturally, Granite 4.2 uses a decoder-only dense transformer with grouped-query attention, rotary position embeddings, SwiGLU MLPs, and RMSNorm. The 8B and 30B models also undergo an agentic RL stage in real sandboxed environments, learning to call tools, edit and run code, drive a terminal, and search the web. Every model supports native tool calling and includes a thinking/non-thinking switch plus a low-effort thinking mode for easy questions. The models can be served through OpenAI-compatible endpoints, for example with vLLM, and are supported in SGLang.

The second post introduces granite-speech-5.0-470m-turboctc and granite-speech-5.0-470m-turboctc-nc, two 470M-parameter English speech recognition models. The team reports throughput above 12,600 RTFx on an NVIDIA H200 GPU, meaning batched inference can transcribe more than 3.5 hours of speech per second. The encoder-only design is more than 20 times faster than previous Granite Speech models and is positioned for speech-to-text on edge devices, though it drops capabilities such as speech translation and keyword biasing.

The two speech models differ in training data and licensing. The noncommercial version, granite-speech-5.0-470m-turboctc-nc, trains on additional data and carries a CC-BY-NC-SA-4.0 license; the other model is Apache 2.0-licensed. According to unofficial results on public English short-form test sets from the OpenASR Leaderboard, the noncommercial model scores an aggregate 4.85% word error rate and the Apache 2.0 model scores 5.00%. On the FFASR Leaderboard as of Aug. 25, the two models ranked ninth and fifth in far-field accuracy while being the fastest two models, with official results.

Both speech models are built from 16 Conformer blocks, use chunkwise attention, and optimize CTC loss. They operate at 12.5 tokens per second, down from the previous 50 characters per second, using three stages of 2x subsampling after the log Mel spectrogram front end. The noncommercial model uses SentencePiece tokenization; the Apache 2.0 model uses BPE tokenization, with tokenizers trained on speech transcripts. A WebGPU demo of streaming speech recognition is available, but it requires Chrome or Edge browsers.