IBM Granite 4.2: Apache 2.0 3B/8B/30B with agentic RL and 128K context

IBM shipped Granite 4.2 on August 26 and it is the most quietly useful open release of the month. Three sizes, 3B, 8B and 30B, all Apache 2.0, all with native 128K context. The 8B and 30B variants received specialized reinforcement learning for agent tasks: tool use, terminal work, search and multi-step execution. Distribution is everywhere you would want it on day one: Hugging Face, Ollama, GitHub, LM Studio and the usual inference providers.

The enterprise split

The interesting structure in the open-model market this summer is a split. At one end, models like GLM-5.3-Flash, released the same day at 320B total and 18B active with a 1M context, chasing frontier coding quality at a fraction of API cost. At the other, small, controllable models with native tool calling that a company can run entirely inside its own network. Granite 4.2 is squarely the second kind, and IBM is explicit about the target: local enterprise agents over sensitive data with a predictable cost.

That maps to real constraints. Customer data that cannot leave the building. A per-seat budget that a metered frontier API blows through the first time someone loops an agent. A compliance officer who wants to know exactly which weights are running and under which license. Apache 2.0 answers the last question cleanly, which is more than most community licenses do.

The question for a local agent is not how smart the model is. It is whether it calls the right tool with the right arguments on the twentieth step, every time.

Why the RL matters more than the size

A 3B model that answers questions is a commodity. An 8B model that reliably emits a well-formed tool call, reads the result, and decides the next step is a different product, and that reliability is what IBM says the agentic RL targets. Native 128K context means an agent can hold a real repository slice or a long tool transcript without a retrieval layer in front of it. On Apple Silicon, an 8B model at 4-bit sits comfortably in the unified memory of a mid-range machine, and a 30B fits on a Max-class configuration with room for the KV cache.

The function-calling plumbing on the Ollama side is in Ollama function calling, and the wider picture in local models. Granite 4.2 is the candidate you try when a commercial-use license requirement rules out half the list.

What to benchmark on an M-series box

The obvious comparison is against Qwen3.8 and GLM-5.3-Flash on MLX. Here is what I would measure, in order.

• Tool-call validity rate. Over a fixed set of 100 tasks with a known tool schema, how often is the call parseable and correctly typed. This matters more than any leaderboard.
• Multi-step completion. Tasks that need three to five tool calls in sequence. Count full completions, not partial credit.
• Context behaviour at 64K and 128K. Prefill time, decode tokens per second, and whether quality holds when the relevant fact sits at the start of a long transcript.
• Memory footprint per quantization. 4-bit and 8-bit for the 8B and 30B, with KV cache at your real context length.
• Cost per completed task. Wall time times power on the Mac versus the equivalent frontier API call. This is the number the budget owner wants.

Run the same suite on a comparably sized Qwen3.8 and on GLM-5.3-Flash if you have the memory. The winner for local agents is whichever one fails least, not whichever scores highest.

Where this fits and where it does not

Granite 4.2 fits internal agents over sensitive data, fixed tool sets, and teams that need a license a lawyer will sign off in an afternoon. It does not fit open-ended coding across unfamiliar repos, anything where a frontier model's reasoning is the product, or teams without the discipline to build the suite above.

The honest limitation

IBM's announcement, as it reached me, carries no benchmark numbers, and I am not going to invent them. "Specialized RL for agent tasks" describes a training recipe, not a measured result, and until the harness above has run on real hardware I cannot tell you whether the 8B beats a comparable Qwen at tool calling or loses to it. Apache 2.0 and 128K are facts. Everything about quality is a test you still have to run.