News Hub

Polymorphic Inference self-tunes open model inference as models, traffic, and hardware change.

Together Compiler optimizes open models using agents and persistent engines, delivering up to 3.7x higher throughput.

Together AI offers post-training service for open LLMs with custom SFT or RL on managed infrastructure.

Together AI, IBM Cloud, and NVIDIA built a B300 inference cluster for production-scale open models.

Together Link integrates open source models like GLM 5.3 and Kimi K3 into existing coding agents, reducing model costs by over 50 percent.

Together AI released Tev1-4B-experimental, a Jailbreak-detection classifier based on Qwen3.5 4B, available for fine-tuning on their serverless platform.

Together AI enables canary rollouts for model upgrades using staged traffic ramps, metric gates, and automatic rollback.

A global bank used Together AI's Dedicated Model Inference to give engineering teams direct control over scaling and testing coding agents.

Together AI published a five-stage playbook for migrating from closed to open source models.

Together AI adds more models, live metrics, and controls to its fine-tuning service.

ThunderKittens ported to NVIDIA Vera Rubin NVL72 achieves 22 PFLOPS, competitive with cuBLAS.

Together AI offers preemptible GPU compute at 50% of on-demand rates with five-minute drain windows.

Together AI describes the open source AI stack.

GLM-5.3 Flash achieves 17x lower cost than GLM-5.3 with only 5.6 point pass@1 reduction on DeepSWE.

GLM-5.3 achieves better pass@4 at half the cost of GPT-5.6 Sol on DeepSWE.

GLM-5.3 costs 5.4x less than Claude Fable 5 while matching pass@1 on DeepSWE.

DeepSeek V4 Pro 0813 achieved 82.7% pass rate on DeepSWE cascade routing, outperforming Claude Fable 5 on cost-adjusted coding tasks.

GPT-5.6 Luna leads DeepSeek-V4 Flash by 14 points on DeepSWE; DeepSeek delivers 4.8x solves per dollar.

Together AI provides guidance on autoscaling metrics, scale windows, and cold start budgeting for LLM inference endpoints.

ThunderAgent scheduler achieves over 2x single-node throughput and near-linear multi-node scaling for agentic inference by eliminating KV cache thrashing.