Together Research says LLMs can patch faulty database query plans instead of regenerating them from scratch, and claims up to 4.78x speedups on some TPC-H and TPC-DS workloads. The tweet points to DBPlanBench, a DataFusion-based harness that exposes a physical operator graph to an LLM and uses iterative search to refine plan edits.
Together AI said on April 3, 2026 that Wan 2.7 from Alibaba Cloud is now available on its platform. The accompanying product post says text-to-video is live now, with image-to-video, reference-to-video, and video edit workflows rolling out on the same API, auth, and billing surface.
Together Research said on March 31, 2026 that Aurora is an open-source framework for adaptive speculative decoding that learns from live inference traces and updates the speculator asynchronously without interrupting serving. Together’s blog and paper say Aurora reframes the problem as asynchronous RL and can deliver 1.25x additional speedup over a strong static speculator as traffic shifts.
Together Research said on March 27, 2026 that a smaller model using divide-and-conquer can match or outperform GPT-4o on long-context tasks, with the work accepted at ICLR 2026. Together's blog and the arXiv paper say the method uses a planner-worker-manager pipeline and explains long-context failures in terms of task, model, and aggregator noise.
Together AI said on March 19, 2026 that its fine-tuning service now supports tool-call, reasoning, and vision-language workflows. The linked Together AI blog adds 100B+ parameter model support, datasets up to 100GB, up to 6x higher throughput on large MoE models, and upfront cost plus ETA estimates.
Together AI said on March 19, 2026 that its fine-tuning service now supports tool calling, reasoning, and vision-language model training, with up to 6x higher throughput on MoE architectures. The company says the update also targets very large models, supports datasets up to 100GB, and adds pre-run cost estimates plus live ETAs during training.
Together AI and collaborators introduced Mamba-3 as an inference-first state space model. Hacker News traction centered on faster prefill+decode latency, a stronger recurrence design, and open-sourced high-performance kernels.
A high-score Hacker News discussion surfaced Together AI's CDLM post, which claims up to 14.5x latency improvements for diffusion language models by combining trajectory-consistent step reduction with exact block-wise KV caching.