Industry Updates

When Vercel Agent works on a production problem, it investigates automatically under read-only access, then proposes a plan listing the tasks it would do and the permissions they need, and only after the user approves that plan does it receive access beyond the read-only baseline. Access is defined by scopes, the named capabilities recognized by Vercel's authorization system; a plan requesting more than the baseline, or including supported external actions, waits for approval as a whole, and the runtime checks the plan is active before selecting its approved credential while Vercel APIs check each request against that credential's permissions.

When Vercel Agent works on a production problem, it investigates automatically under read-only access, then proposes a plan listing the tasks it would do and the permissions they need, and only after the user approves that plan does it receive access beyond the read-only baseline. Access is defined by scopes, the named capabilities recognized by Vercel's authorization system; a plan requesting more than the baseline, or including supported external actions, waits for approval as a whole, and the runtime checks the plan is active before selecting its approved credential while Vercel APIs check each request against that credential's permissions.

Moonshot released the reasoning model Kimi K3.1 Pro, which scores 55 on the Artificial Analysis Intelligence Index and places fourth, 11 points above Kimi K3's maximum of 44. Its largest gains are agentic: 1,846 Elo on GDPval-AA, 312 Elo above Kimi K3, and the highest score of any model on Terminal-Bench 4.0. Its weights are not yet published; Moonshot is expected to release them soon after launch, at which point it would be the highest-scoring open-weights model and the gap to proprietary models on the index would narrow from 12 points to 3.

Moonshot released the reasoning model Kimi K3.1 Pro, which scores 55 on the Artificial Analysis Intelligence Index and places fourth, 11 points above Kimi K3's maximum of 44. Its largest gains are agentic: 1,846 Elo on GDPval-AA, 312 Elo above Kimi K3, and the highest score of any model on Terminal-Bench 4.0. Its weights are not yet published; Moonshot is expected to release them soon after launch, at which point it would be the highest-scoring open-weights model and the gap to proprietary models on the index would narrow from 12 points to 3.

Moonshot AI released Kimi K3.1 Pro, the first model in its K3.1 series and its most capable 3T-class open-weight model to date, with a 1-million-token context window, alongside a Highspeed variant delivering 2-3x the speed. It is a post-trained version of Kimi K3 built with substantially larger post-training compute, supported by optimizations across the RL framework, long-context inference, sandbox infrastructure and rollout scheduling. On Terminal-Bench 4.0, which tests coding agents on realistic terminal tasks, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%.

Moonshot AI released Kimi K3.1 Pro, the first model in its K3.1 series and its most capable 3T-class open-weight model to date, with a 1-million-token context window, alongside a Highspeed variant delivering 2-3x the speed. It is a post-trained version of Kimi K3 built with substantially larger post-training compute, supported by optimizations across the RL framework, long-context inference, sandbox infrastructure and rollout scheduling. On Terminal-Bench 4.0, which tests coding agents on realistic terminal tasks, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra at 58.2% and narrowly behind Claude Opus 5.5 at 66.4%.

Anthropic has published the safety design behind Claude Duet, a personal agent still in research preview that uses a cloud computer to read inboxes, draft replies, book things and run routines after the app is closed. Before sending a potentially sensitive message, paying or deleting anything, it shows the action, destination and every recipient and waits for approval. Passwords, browser cookies and sign-in tokens never enter the model, and a safety classifier separate from the main model checks outgoing messages for sensitive information. Anthropic says these safeguards will still occasionally miss or over-flag, so they run alongside the approval mechanisms.

Anthropic has published the safety design behind Claude Duet, a personal agent still in research preview that uses a cloud computer to read inboxes, draft replies, book things and run routines after the app is closed. Before sending a potentially sensitive message, paying or deleting anything, it shows the action, destination and every recipient and waits for approval. Passwords, browser cookies and sign-in tokens never enter the model, and a safety classifier separate from the main model checks outgoing messages for sensitive information. Anthropic says these safeguards will still occasionally miss or over-flag, so they run alongside the approval mechanisms.

NVIDIA's Global Performance Tuner, released in TensorRT 11.1, represents compiler scheduling choices as searchable build routes, a concrete combination of compiler knob values, and selects the fastest measured end-to-end engine among those that meet a user-supplied accuracy boundary. On ten evaluated workloads, with each default build route normalized to 1.0x, the best eligible route found during tuning delivered 1.14x to 1.63x. Accuracy decides which candidates are acceptable and performance ranks only those, so a tuned route is tied to its model, input shapes, target GPU and software versions.

NVIDIA's Global Performance Tuner, released in TensorRT 11.1, represents compiler scheduling choices as searchable build routes, a concrete combination of compiler knob values, and selects the fastest measured end-to-end engine among those that meet a user-supplied accuracy boundary. On ten evaluated workloads, with each default build route normalized to 1.0x, the best eligible route found during tuning delivered 1.14x to 1.63x. Accuracy decides which candidates are acceptable and performance ranks only those, so a tuned route is tied to its model, input shapes, target GPU and software versions.

NVIDIA released OpenShell 0.1.x, a runtime that lets fleets of autonomous AI agents read files, install packages, call APIs and use credentials without unrestricted access to data, secrets or network. Each agent runs in an isolated sandbox where kernel controls confine file access and system calls, every network connection passes a policy check before leaving, and agents never see real credentials. Before a policy change is approved, formal verification flags risky new access it would grant, such as reaching a new host with credentials, and those changes wait for human review.

NVIDIA released OpenShell 0.1.x, a runtime that lets fleets of autonomous AI agents read files, install packages, call APIs and use credentials without unrestricted access to data, secrets or network. Each agent runs in an isolated sandbox where kernel controls confine file access and system calls, every network connection passes a policy check before leaving, and agents never see real credentials. Before a policy change is approved, formal verification flags risky new access it would grant, such as reaching a new host with credentials, and those changes wait for human review.

Amazon released TorchNeuron, an open-source backend that translates PyTorch code into operations executable on AWS Trainium, so PyTorch code already running on other accelerators runs on Trainium with minimal or no code changes. An optional adaptive eager mode groups operations into small graphs that are compiled and cached, approaching fully compiled performance while keeping eager mode's debugging experience. For evaluation, AI agents swept Hugging Face repositories to identify 841 representative models covering the range of architectures, and tens of thousands of tensor shapes were distilled for tuning and benchmarking.

Amazon released TorchNeuron, an open-source backend that translates PyTorch code into operations executable on AWS Trainium, so PyTorch code already running on other accelerators runs on Trainium with minimal or no code changes. An optional adaptive eager mode groups operations into small graphs that are compiled and cached, approaching fully compiled performance while keeping eager mode's debugging experience. For evaluation, AI agents swept Hugging Face repositories to identify 841 representative models covering the range of architectures, and tens of thousands of tensor shapes were distilled for tuning and benchmarking.

MiniMax open-sourced MiniMax-M3.1-Flash, and vLLM supports it on day one on NVIDIA Blackwell, NVIDIA Rubin and AMD MI350-series GPUs. The model is quantization-aware trained for an NVFP4 KV cache and stores its routed MoE weights in NVFP4, shrinking the checkpoint from 451.5 GB to 250.1 GB and the KV cache from 61,440 to 34,560 bytes per token. On four NVIDIA B300 GPUs with DEP4 deployment and 256K-token sequences, M3.1 fits 76 concurrent sequences where M3 with an FP8 KV cache fits 28.

MiniMax open-sourced MiniMax-M3.1-Flash, and vLLM supports it on day one on NVIDIA Blackwell, NVIDIA Rubin and AMD MI350-series GPUs. The model is quantization-aware trained for an NVFP4 KV cache and stores its routed MoE weights in NVFP4, shrinking the checkpoint from 451.5 GB to 250.1 GB and the KV cache from 61,440 to 34,560 bytes per token. On four NVIDIA B300 GPUs with DEP4 deployment and 256K-token sequences, M3.1 fits 76 concurrent sequences where M3 with an FP8 KV cache fits 28.

IBM Research released Granite Timeseries Ensemble, a default forecasting recipe that combines four open Granite time-series models with different architectures and returns median and quantile forecasts without retraining or fitting ensemble weights. Previously a user had to judge which single model suited a given series. Across 97 GIFT-Eval configurations the probabilistic ensemble reaches a MASE of 0.683 against 0.685 for the strongest member, PatchTST-FM-r2, a small margin whose main benefit is not having to pick a model in advance.

IBM Research released Granite Timeseries Ensemble, a default forecasting recipe that combines four open Granite time-series models with different architectures and returns median and quantile forecasts without retraining or fitting ensemble weights. Previously a user had to judge which single model suited a given series. Across 97 GIFT-Eval configurations the probabilistic ensemble reaches a MASE of 0.683 against 0.685 for the strongest member, PatchTST-FM-r2, a small margin whose main benefit is not having to pick a model in advance.

TorchTitan's annual update adds PyTorch-native training support for open-weight models including Muse Spark Open, Qwen3.5, Kimi K3 and DeepSeek-V4, and introduces TitanRL, which keeps one model from pretraining through post-training to inside vLLM so training and inference no longer run two copies that can quietly disagree. In supported on-policy configurations the trainer and generator produce bitwise identical log probabilities, leaving exactly zero gap at every step. The update also reports DeepSeek V3 671B throughput on the MLPerf training benchmark and infrastructure work around FlexShard and torch_checkpointing.

TorchTitan's annual update adds PyTorch-native training support for open-weight models including Muse Spark Open, Qwen3.5, Kimi K3 and DeepSeek-V4, and introduces TitanRL, which keeps one model from pretraining through post-training to inside vLLM so training and inference no longer run two copies that can quietly disagree. In supported on-policy configurations the trainer and generator produce bitwise identical log probabilities, leaving exactly zero gap at every step. The update also reports DeepSeek V3 671B throughput on the MLPerf training benchmark and infrastructure work around FlexShard and torch_checkpointing.