Modal Account Context
What happened
- Modal has no direct news, but briefings flag TPU 8i/8t and Kimi 2.6 momentum as relevant deployment context. - The specific signals are Google’s TPU split and Kimi 2.6’s day‑zero integrations with platforms like Baseten. - Those signals could change Modal’s cloud versus on‑prem inference tradeoffs around latency, token cost, and tooling choices. (blog.google) (softmaxdata.com)
Why it matters
Modal did not announce a product this week, but two April 2026 releases changed the backdrop for anyone deciding where to run AI workloads. Google split its eighth-generation Tensor Processing Units into TPU 8t for training and TPU 8i for inference on April 22, while Moonshot AI shipped Kimi K2.6 with immediate availability across its own API and third-party platforms including Baseten. (blog.google) (kimi.com) (baseten.co) A chip is the engine under an AI service, and Google is now selling two different engines for two different jobs. The company said TPU 8t is built for large-scale training, while TPU 8i is tuned for low-latency inference and will reach general availability later in 2026. (blog.google) (cloud.google.com) Google said the split reflects a change in how AI systems are built. In its technical write-up, the company said pre-training, post-training, and real-time serving now have different bottlenecks, and it sized TPU 8t around 9,600 chips in one superpod while positioning TPU 8i for large-scale inference and reinforcement learning. (cloud.google.com) A model is the software brain, and Kimi K2.6 is one of the new open models pushing harder on long-running agent work. Moonshot AI said on April 21 that K2.6 is open sourced, supports text, image, and video input, and is available through Kimi.com, the Kimi app, the API, and Kimi Code. (kimi.com) (platform.kimi.ai) Moonshot AI said K2.6 supports 256,000 tokens of context and multi-step tool use, which are the features developers lean on when models read long codebases or call outside services. In one company example, K2.6 made more than 4,000 tool calls over 12 hours and raised throughput from about 15 tokens per second to about 193. (platform.kimi.ai) (kimi.com) Baseten listed Kimi K2.6 in its model library within days of the release, and its documentation says model APIs use OpenAI-compatible endpoints on shared infrastructure managed by Baseten. That is the kind of day-zero packaging that can shorten the path from a new model launch to a production deployment. (baseten.co) (docs.baseten.co) Modal’s pitch sits on the other side of that decision tree. Its homepage says developers can run inference, training, and batch processing with sub-second cold starts and instant autoscaling, and a company blog post last week argued that agent workloads swing between one GPU, dozens of parallel jobs, and multi-GPU clusters in the same session. (modal.com 1) (modal.com 2) That leaves three concrete variables in play for platforms like Modal: latency, token cost, and tooling. If Google is carving out custom silicon for fast serving and vendors like Baseten are wrapping new open models behind managed APIs on day one, cloud users get more reasons to compare a serverless GPU platform with a managed model endpoint or a dedicated in-house stack. (blog.google) (docs.baseten.co) (modal.com) The near-term question is not whether Modal changed this week; it is whether the market around Modal did. Google’s TPU 8i and 8t split, and Kimi K2.6’s fast spread into hosted inference catalogs, added new reference points for how developers price speed, flexibility, and control in 2026. (cloud.google.com) (baseten.co)
Key numbers
- Modal has no direct news, but briefings flag TPU 8i/8t and Kimi 2.6 momentum as relevant deployment context.
- The specific signals are Google’s TPU split and Kimi 2.6’s day‑zero integrations with platforms like Baseten.
- (blog.google) (softmaxdata.com) Modal did not announce a product this week, but two April 2026 releases changed the backdrop for anyone deciding where to run AI workloads.
- Google split its eighth-generation Tensor Processing Units into TPU 8t for training and TPU 8i for inference on April 22, while Moonshot AI shipped Kimi K2.6 with immediate availability across its own API and third-party platforms including Baseten.
What happens next
- The company said TPU 8t is built for large-scale training, while TPU 8i is tuned for low-latency inference and will reach general availability later in 2026.
- That is the kind of day-zero packaging that can shorten the path from a new model launch to a production deployment.
- Those signals could change Modal’s cloud versus on‑prem inference tradeoffs around latency, token cost, and tooling choices.
Quick answers
What happened in Modal Account Context?
Modal has no direct news, but briefings flag TPU 8i/8t and Kimi 2.6 momentum as relevant deployment context. The specific signals are Google’s TPU split and Kimi 2.6’s day‑zero integrations with platforms like Baseten. Those signals could change Modal’s cloud versus on‑prem inference tradeoffs around latency, token cost, and tooling choices. (blog.google) (softmaxdata.com)
Why does Modal Account Context matter?
Modal did not announce a product this week, but two April 2026 releases changed the backdrop for anyone deciding where to run AI workloads. Google split its eighth-generation Tensor Processing Units into TPU 8t for training and TPU 8i for inference on April 22, while Moonshot AI shipped Kimi K2.6 with immediate availability across its own API and third-party platforms including Baseten. (blog.google) (kimi.com) (baseten.co) A chip is the engine under an AI service, and Google is now selling two different engines for two different jobs. The company said TPU 8t is built for large-scale training, while TPU 8i is tuned for low-latency inference and will reach general availability later in 2026. (blog.google) (cloud.google.com) Google said the split reflects a change in how AI systems are built. In its technical write-up, the company said pre-training, post-training, and real-time serving now have different bottlenecks, and it sized TPU 8t around 9,600 chips in one superpod while positioning TPU 8i for large-scale inference and reinforcement learning. (cloud.google.com) A model is the software brain, and Kimi K2.6 is one of the new open models pushing harder on long-running agent work. Moonshot AI said on April 21 that K2.6 is open sourced, supports text, image, and video input, and is available through Kimi.com, the Kimi app, the API, and Kimi Code. (kimi.com) (platform.kimi.ai) Moonshot AI said K2.6 supports 256,000 tokens of context and multi-step tool use, which are the features developers lean on when models read long codebases or call outside services. In one company example, K2.6 made more than 4,000 tool calls over 12 hours and raised throughput from about 15 tokens per second to about 193. (platform.kimi.ai) (kimi.com) Baseten listed Kimi K2.6 in its model library within days of the release, and its documentation says model APIs use OpenAI-compatible endpoints on shared infrastructure managed by Baseten. That is the kind of day-zero packaging that can shorten the path from a new model launch to a production deployment. (baseten.co) (docs.baseten.co) Modal’s pitch sits on the other side of that decision tree. Its homepage says developers can run inference, training, and batch processing with sub-second cold starts and instant autoscaling, and a company blog post last week argued that agent workloads swing between one GPU, dozens of parallel jobs, and multi-GPU clusters in the same session. (modal.com 1) (modal.com 2) That leaves three concrete variables in play for platforms like Modal: latency, token cost, and tooling. If Google is carving out custom silicon for fast serving and vendors like Baseten are wrapping new open models behind managed APIs on day one, cloud users get more reasons to compare a serverless GPU platform with a managed model endpoint or a dedicated in-house stack. (blog.google) (docs.baseten.co) (modal.com) The near-term question is not whether Modal changed this week; it is whether the market around Modal did. Google’s TPU 8i and 8t split, and Kimi K2.6’s fast spread into hosted inference catalogs, added new reference points for how developers price speed, flexibility, and control in 2026. (cloud.google.com) (baseten.co)
