IBM's open Granite 4.2 models learn agent skills via reinforcement learning
A string of open-weight releases this week keeps pushing what small, deployable models can do. IBM is the latest, rolling out a reasoning-tuned Granite family that trains the bigger variants to actually use tools and code in real sandboxes before shipping them under a permissive Apache 2.0 license.
IBM released Granite 4.2, a family of dense, open-weight reasoning models in 3B, 8B, and 30B sizes, its first to be post-trained with reinforcement learning that teaches them agent-style behavior. Each model was pre-trained from scratch on roughly 15 trillion tokens with a context window stretched to 512K tokens, then fine-tuned on chain-of-thought and agentic-trajectory data before going through a multi-stage RL pipeline. The 8B and 30B models get the headline treatment: an "agentic RL" block that trains them to call tools, edit and run code, drive a terminal, and search the web inside real (not simulated) environments, rewarded on whether the task is actually solved. Every model supports a thinking/non-thinking switch, a low-effort mode, native tool calling, and quantization to FP8, FP4, and GGUF for deployment.
The significance is two-fold. First, it's the clearest sign yet that reinforcement learning to build agents is becoming the standard post-training recipe for open models — IBM trained with NVIDIA's NeMo-RL on a CoreWeave-hosted GB200 cluster, the same RL-first playbook Anthropic-style labs use for Claude. Second, it's aimed squarely at enterprises that want frontier-ish agentic behavior without handing code and internal data to a hosted API; the 30B model posts a 57.00 on SWE-Bench Verified and 89.96 on RULER at 64K context, respectable numbers for open weights its size. The catch is the oldest one for open reasoning models: verifiable-reward scores look strong, but real-world agentic reliability is still an open question until people run it.
Nvidia doubled compute for entry-level edge robotics with the Jetson Orin Nano 2, a small AI "brain" for drones and lightweight machines. The module jumps to 78 trillion operations per second of AI compute, 8 gigabytes of memory, and an eight-core CPU while drawing 40 percent less power than its predecessor in its 15-watt mode — enough, Nvidia says, to run vision and perception tasks on-device rather than relaying to the cloud, which lets a robot react in the moment instead of waiting on a round-trip. Partners including Wing Aviation, Cognex, Doosan Bobcat, and Matic Robotics are already building on it; Wing plans to use it for real-time perception on its drone-delivery fleet. It's another sign of the pattern we flagged with earlier edge silicon: as small and medium models close in on last year's flagship accuracy, the compute that can run them keeps shrinking toward the device. The module and developer kit go generally available in the first half of 2027.
Cisco expanded its rack-scale Secure AI Factory with Nvidia for neocloud and sovereign-cloud deployments, in a partnership with Super Micro. The portfolio pairs Cisco's Silicon One switches on the front end with Nvidia Spectrum-X networking on the back end, unified by the Nexus One management platform, with liquid-cooled N9000 switches built to handle racks that can exceed 200 kilowatts. Cisco frames it as a play to let enterprises treat an AI factory like a bolt-on expansion of existing infrastructure rather than a greenfield overhaul, with Validated Services to certify that builds match reference architectures. It's an infrastructure story, but it matters for the AI story: compute, networking, and cooling are the parts of the build-out vendors keep racing to industrialize as agentic inference pushes beyond GPU count into how you move data around at scale. The Super Micro compute rolls out as part of the line starting in October.
What to watch: IBM's Granite 4.2 is open and on Hugging Face today — the real test is how the agentic-RL promise holds up in the wild once developers push the 30B past its benchmark suite.
Reinforcement learning is now the standard way to teach open models to act with tools — does that close the gap on frontier agents, or just widen the gap between benchmark scores and real-world reliability? Tell us in the comments.
Sources: Granite 4.2 on GitHub
