IBM Releases Granite 4.2: Bringing Native Reasoning and Agentic RL to Open Enterprise Models
IBM has launched Granite 4.2, a household of open reasoning language fashions in 3B, 8B, and 30B parameter sizes. Unlike earlier Granite releases, which have been instruction-following assistants, Granite 4.2 is constructed round specific reasoning. Every mannequin can emit a sequence of thought earlier than answering, and each mannequin exposes a considering / non-thinking swap plus a low-effort mode that spends a brief reasoning finances on simple questions. The fashions are decoder-only dense transformers, pre-trained from scratch on roughly 15 trillion tokens, then post-trained via a multi-stage reinforcement studying chain. For the 8B and 30B, that chain consists of an agentic RL block the place the mannequin learns to edit code, drive a terminal, and run net searches inside actual sandboxed environments. All three ship beneath Apache 2.0. IBM additionally launched two 470M-parameter Granite Speech 5.0 Turbo CTC fashions alongside the LLMs.
Is it deployable?
Yes, All three Granite 4.2 language fashions ship beneath Apache 2.0, so obtain, fine-tuning, and business manufacturing use carry no licensing gate.
- Which firms: The 3B matches solo builders and startups working on a laptop computer via Ollama or LM Studio, particularly with the launched GGUF quants down to Q4_K_M. The 8B fits mid-market groups on a single fashionable GPU. The 30B targets enterprises with A100/H100-class capability, or FP8/NVFP4 serving on vLLM. Regulated organizations get the extra good thing about on-prem weights.
- Industries: Software and developer tooling, monetary companies, healthcare, telecom, public sector, and contact facilities, which is the place the brand new speech fashions land.
- Applications: Software engineering brokers, terminal and DevOps automation, deep-research and search brokers, long-document RAG, structured software calling, and high-volume transcription.
Architecture
Granite 4.2 is a decoder-only dense transformer, not a hybrid or MoE design. Core elements are Grouped Query Attention with 8 KV heads, RoPE with θ = 10,000,000, SwiGLU MLPs, RMSNorm (ε = 1e-5), untied enter/output embeddings, and bfloat16 precision.
The 3B makes use of 40 layers at embedding measurement 2560. The 8B makes use of 40 layers at 4096. The 30B goes to 64 layers with an MLP hidden measurement of 32,768. The printed structure desk lists a 131,072-token (128K) sequence size, whereas the five-phase pre-training run features a long-context section extending to 512K tokens. Pre-training covers roughly 15 trillion tokens from scratch.
The coaching pipeline is the precise story
Supervised fine-tuning makes use of about 7.2 million samples, roughly 100B tokens with ~65B trainable. The combination is 31.6% agentic and 68.4% non-agentic, and software program engineering is 69% of the agentic slice. Trajectories have been generated throughout harnesses together with OpenPalms, SWE-agent, Terminus-2, MiniSWE, Codex, and Goose. Quality management used GPT-OSS-120B and Gemma 4 as judges, plus SHA-256 deduplication over the instruments and messages fields.
Post-training is a multi-stage, multi-environment RL chain, not a single cross. Each stage is a separate asynchronous GRPO run that warm-starts from the earlier checkpoint, with a leave-one-out baseline as a substitute of a price community and truncated significance sampling to sure off-policy drift. The order is RLVR, then ability boosters, then SWE, Terminal, Search, then RLHF.
The agentic RL block runs solely on the 8B and 30B. The 3B takes foundational RL and alignment solely. That single design selection explains a lot of the functionality hole throughout sizes. Training ran on NeMo-RL and NeMo-Gym over an NVIDIA GB200 NVL72 cluster hosted by CoreWeave.
Two supporting items matter: 1 trillion tokens of artificial code from IBM’s CodeAlchemy pipeline, and a speculative decoding layer for sooner serving.
Reported outcomes
IBM’s numbers, by measurement (3B / 8B / 30B):
| Benchmark | 3B | 8B | 30B |
|---|---|---|---|
| SWE-Bench Verified | NA | 47.67 | 57.00 |
| Terminal-Bench 2.1 | NA | 20.56 | 29.24 |
| τ³-bench | 50.99 | 66.34 | 68.05 |
| BFCL (v4) | 52.41 | 50.29 | 61.39 |
| AIME25 | 78.33 | 86.67 | 89.17 |
| GPQA | 54.80 | 64.14 | 66.41 |
| MMLU-Pro | 67.84 | 74.04 | 77.60 |
| RULER 128K | 55.30 | 71.41 | 81.38 |
Speech: 470M parameters, no LLM spine
The Turbo CTC fashions are available at 470 million parameters and drop the LLM spine totally, utilizing connectionist temporal classification to map audio to textual content. IBM stories an RTFx throughput close to 12,600 on a single H200, in opposition to roughly 6,000 for present pace leaders on the Open ASR leaderboard. A WebGPU demo is dwell.
Key Takeaways
- Granite 4.2 ships dense 3B/8B/30B reasoning fashions beneath Apache 2.0.
- A considering / low-effort / non-thinking swap is uncovered within the chat template.
- Agentic RL (SWE, Terminal, Search) trains solely the 8B and 30B.
- The 30B hits 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1.
- Granite Speech 5.0 Turbo CTC is 470M parameters with no LLM spine.
Check out the IBM Research blog, the technical write-up, and the GitHub repo. Also, be at liberty to observe us on Twitter and don’t overlook to be part of our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
The submit IBM Releases Granite 4.2: Bringing Native Reasoning and Agentic RL to Open Enterprise Models appeared first on MarkTechPost.
