The Sizes Are Cut to Match Deployment Targets
The three sizes map to three machines rather than three performance tiers. The 3B targets laptops through Ollama or LM Studio, with GGUF quantizations down to Q4_K_M. The 8B fits a single modern GPU. The 30B needs A100/H100-class capacity, or FP8/NVFP4 serving on vLLM. On the tooling side there is OpenAI-compatible function calling, plus vLLM and SGLang support. Inference control comes in three settings rather than two: thinking, non-thinking, and a low-effort mode that burns fewer reasoning tokens on easy questions. For anyone paying per token or running locally, that switch matters more than a benchmark score — it directly determines what the same model actually costs you.
Only the 8B and 30B Got Agentic Training
One distinction worth keeping straight: foundational reinforcement learning covers all three sizes, but the agentic RL for SWE, terminal and search was applied only to the 8B and 30B. Those two learned to use tools, write and run code, and search the web inside real sandbox environments. If you are after agent capability, the 3B did not receive the same training. The published results put the 30B at 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1. IBM frames this generation as reasoning-focused, purpose-built for the agentic workflows enterprise use cases require. Shipping alongside were two 470M-parameter Granite Speech 5.0 Turbo CTC models, which IBM says reach RTFx throughput near 12,600 on a single H200.
The Real Differentiators Are the License and the Recipe
Apache 2.0 means you can download, fine-tune and put it into production with no additional restrictions — a hard requirement in on-premises enterprise deployments, and where plenty of comparable open-weight models fall down. Rarer still, IBM published the full training recipe, data-mixture proportions and per-stage hyperparameters with the weights, which makes reproduction and continued training possible rather than just handing over a checkpoint. One discrepancy to flag: reported context length varies across coverage, with some outlets citing 128K and others 512K (the ceiling reached through multi-stage pre-training extension). IBM's own blog post does not state the figure directly, so check the model card before making a selection decision rather than copying a secondhand number. For teams choosing a local model, the reason to shortlist this generation is not any single benchmark but the combination: a clean license, real agentic training, and three sizes spanning laptop to server. Not many open models satisfy all three at once.
via: IBM Research blog, The Decoder, The New Stack, MarkTechPost