IBM (Granite)

Thesis

IBM's Granite program is building an open-weights, enterprise-first stack — dense, decoder-only reasoning LLMs plus speech, vision, time-series, and safety models — and distributing it broadly (Hugging Face, Ollama/GGUF, Apple MLX, quantized formats) rather than chasing a single flagship benchmark. The signals point to a deployment-breadth plus agentic-enterprise strategy: native reasoning, tool calling, and RAG primitives in Granite 4.2 W1W2P25, a parallel edge/time-series portfolio P26P17, and governance tooling (Granite Guardian, Granite.Trust policy schema) E15W4. IBM pairs this open stack with commercial distribution (OpenAI and Together AI partnerships) and adjacent quantum/HPC bets E57E58P24. Evidence is thin on hiring and absent on forks, so headcount and upstream-inspection signals remain underdetermined in this pack W5.

Signal desks

Hiring

• One cited signal: IBM Research's Core Foundation Models team recruiting researchers and engineers to work on Granite, an open-source effort spanning agentic AI models and RL platforms W5. No locations, team counts, or data/eval-specific roles are cited.

Forks

• No cited evidence in this pack.

Releases

• Granite 4.2 language family (3B/8B/30B), Apache 2.0, dense decoder-only with switchable thinking mode, tool calling, RAG, structured output P25P27E3E4E6W1W2.
• MLX variants for Apple Silicon (3b/8b/30b bf16) P10P11P12E12E13E14.
• Quantized variants: FP8, NVFP4, MXFP4 P28E42E46E47E48E49E55E56.
• Granite Speech 5.0 TurboCTC (470M edge ASR); earlier 4.1 speech, vision, guardian, and switch models P26E1E5E7E8E9E15E41.
• Granite-TimeSeries PatchTST-FM-r2, #1 zero-shot on GIFT-Eval P17P18E16.
• granite-common utilities: RAG intrinsics, hallucination/citation processing, certainty, gpt-oss support P9P13P14P15.
• GGUF pre-release/test tags for language, speech, and a 120B OSS line E27E40E43E45.

Talking

• Granite 4.2 launch framing: reasoning-focused release, 512K context, agentic/RL training in real environments W1W2W3.
• Governance research: Granite.Trust Policy Tools (YAML Actionable Policy schema; synthetic policy-aligned data) W4.
• Partnerships/distribution: OpenAI (GPT-5.6 into IBM Consulting), Together AI + NVIDIA on IBM Cloud, Apptio AI Value & ROI, free Lightwell for universities/NGOs E57E58E59E60.
• Adjacent PR: K-12 AI readiness + fellowship, US Open AI fan experience + sports study, quantum (HRL acquisition, modular cryogenics, Gordon Bell finalist), dual-architecture mainframe processor P16P19P2P24E44P3E33.

Shipping

Granite 4.2 (3B/8B/30B) is the flagship shipment: dense decoder-only, from-scratch pre-training, native thinking/reasoning, 128K native context (512K extension), tool calling, RAG, and structured JSON, all under Apache 2.0 P25P27W1W2W3. Model cards date the release to Aug 25, 2026 P25P27. Distribution spans Hugging Face (base + FP8/NVFP4/MXFP4), Ollama/GitHub GGUF, and MLX for Apple Silicon P10P28E42E43W1. Non-language shipping is active in parallel: edge ASR (TurboCTC 470M) P26, the time-series foundation model PatchTST-FM-r2 P17, and vision/guardian models E5E15. Supporting utilities ship on a fast cadence: granite-common adds intrinsics, RAG, hallucination/citation processing, and gpt-oss compatibility P9P13P14P15.

Research themes

• Agentic reasoning: 8B/30B trained with reinforcement learning inside real software-engineering, terminal, and web-search environments; a low-effort reasoning mode W3W2.
• RAG intrinsic evaluation: context relevance, answer relevance, hallucination detection, and citation processing in granite-common P9.
• Time-series foundation modeling: Conformer blocks with alternating kernels, overlapping patches, Hamming-window loss, and CauKer synthetic data P17.
• Edge efficiency: 470M CTC ASR and MLX/quantized deployments for laptop/phone-class hardware P26P10E42.
• Governance/safety: GRC evaluations baked into data curation; an Actionable Policy YAML schema with synthetic policy-aligned data P10W4E15.
• Quantum-centric HPC: a 12,635-atom protein simulated with quantum Heron processors plus Fugaku and Miyabi-G supercomputers P3.

Hiring & scaling

Hiring evidence is thin: the pack cites a single secondary-source listing for IBM Research Foundation Models roles (researchers/engineers on Granite, agentic models, RL platforms) W5, with no cited locations, team composition, or data/eval/infrastructure-specific roles. There is no fork evidence to infer upstream dependencies or pre-release inspection. Scaling intent must therefore be read from shipping cadence — rapid granite-common/granite-tsfm releases and dense GGUF test tags — rather than headcount P15E16E43.

Category implications

• Strategy & product: IBM is positioning Granite as the open, enterprise-agentic alternative — native reasoning plus tool/RAG/structured-output, Apache 2.0 for commercial use W1P25W3. Multi-format distribution (Ollama, Apple MLX, NVFP4/MXFP4/FP8) signals deployment-breadth and edge/on-prem focus over a single flagship P10E42E46W1.
• Infrastructure: the Together AI + NVIDIA agreement on IBM Cloud points to scaled open-source inference E58; GGUF test tags reference a 120B OSS model line, suggesting larger-model packaging work E27E28.
• Research: the time-series and speech portfolio diversifies into industrial/ops and audio modalities beyond text P17P26; quantum HPC (HRL acquisition, Gordon Bell finalist) is a parallel research bet P24P3.
• Governance & safety: Granite Guardian and Granite.Trust policy tooling signal an enterprise GRC-driven safety and go-to-market posture E15W4P10.
• GTM: education (K-12 fellowship), sports (US Open), ROI tooling (Apptio AI Value & ROI), and the OpenAI consulting partnership frame commercialization through brand, services, and value-measurement rather than only raw model distribution P16P19E59E57. No revenue figures are cited.

Traction highlights

• Strongest pull on prior-generation Granite 4.1: 30B at 531,761 downloads / 146 likes; speech-4.1-2b at 358,499 downloads E2E1.
• Granite 4.2 (Aug 2026) early traction is modest: 30B 12,037, 3B 22,081, 8B 19,195 downloads E3E4E6; MLX and quantized variants are single-to-double-digit downloads, indicating early/niche adoption E12E13E14E42E46.
• Vision 4.1-4B 114,725, speech-plus 119,929, guardian 32,621 downloads E5E7E15.
• PatchTST-FM-r2 ranks #1 on GIFT-Eval among replicable zero-shot models P17.
• granite-4.2-language-models repo: 25 stars E36.