{"slug": "turbomamba-faster-inference-faster-training-and-better-prediction-with-lexical", "title": "TurboMamba: Faster Inference, Faster Training, and Better Prediction with Lexical Memory", "summary": "TurboMamba, a new language-model architecture combining a global Mamba-3 model with a small Transformer, achieved the lowest mean test prediction loss among twelve small-scale implementations trained on public FineWeb-Edu, according to an arXiv paper (2610.20682v1). TurboMamba's mean negative log-likelihood was 9.4% lower than Mamba-3's, 7.3% lower than the Jamba version's, and 18.3% lower than the Zamba version's, while end-to-end training throughput reached 2.60x, 2.81x, and 4.82x those baselines respectively. On a batch of 32 sequences with 256-token prompts and 128 generated tokens, TurboMamba measured a 3.12x speedup over Mamba-3, 3.53x over the Jamba version, and 6.21x over the Zamba version.", "body_md": "arXiv:2610.20682v1 Announce Type: new \nAbstract: We introduce TurboMamba, a language-model architecture for faster inference, faster training, and better token prediction. A global Mamba-3 model runs once per completed token block. A small Transformer uses this global context and earlier tokens within the block to predict the next token. Three lightweight additions improve local predictions: direct access to the previous block's last token, a small correction to next-token scores, and a learned mixture with recent-token frequencies. We compare twelve small-scale implementations on public FineWeb-Edu, with comparable parameter counts and shared training-budget caps. The baselines include Transformer, Mamba-3, Jamba, and Zamba. The Jamba and Zamba versions use Mamba-3 in place of their original Mamba modules. TurboMamba achieves the lowest mean test prediction loss (negative log-likelihood, NLL) among all twelve models. Its mean NLL is 9.4% lower than that of Mamba-3, 7.3% lower than that of the Jamba version, and 18.3% lower than that of the Zamba version. End-to-end training throughput is 2.60x that of Mamba-3, 2.81x that of the Jamba version, and 4.82x that of the Zamba version. We measure end-to-end batch latency for a batch of 32 sequences, each with a 256-token prompt and 128 generated tokens. On this workload, TurboMamba achieves a 3.12x speedup over Mamba-3, a 3.53x speedup over the Jamba version, and a 6.21x speedup over the Zamba version.", "url": "https://wpnews.pro/news/turbomamba-faster-inference-faster-training-and-better-prediction-with-lexical", "canonical_source": "https://arxiv.org/abs/2610.20682", "published_at": "2026-10-16 04:00:00+00:00", "updated_at": "2026-10-16 04:32:32.604837+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "ai-research", "natural-language-processing"], "entities": ["TurboMamba", "Mamba-3", "Jamba", "Zamba", "Transformer", "FineWeb-Edu", "arXiv"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/turbomamba-faster-inference-faster-training-and-better-prediction-with-lexical", "markdown": "https://wpnews.pro/news/turbomamba-faster-inference-faster-training-and-better-prediction-with-lexical.md", "text": "https://wpnews.pro/news/turbomamba-faster-inference-faster-training-and-better-prediction-with-lexical.txt", "jsonld": "https://wpnews.pro/news/turbomamba-faster-inference-faster-training-and-better-prediction-with-lexical.jsonld"}}