IBM (Granite)

Thesis

IBM is shipping an Apache-2.0 open-model portfolio - dense reasoning text models (Granite 4.2), multilingual speech (Granite Speech 5.0), modular single-endpoint models (Granite Switch 4.1), and time-series forecasting (Granite Timeseries) - while layering an enterprise narrative around agentic-AI governance, sovereignty, and production-scale inference E1E2E3P6P10P19P12P23P9. The composite signal is less a frontier-benchmark play than an open, governable, deploy-anywhere agentic stack aimed at regulated and data-sensitive enterprises, reinforced by Q3 commentary framing clients as deploying 'Super Intelligence' to transform operating models P8.

Signal desks

Hiring

• No cited evidence in this pack. (No open-role, team, location, or job-description signals are provided.)

Forks

• No cited evidence in this pack. (No forked-repo or upstream-dependency signals are provided.)

Releases

• org=ibm-granite · Granite 4.2 dense reasoning family (3B/8B/30B, Apache-2.0, text-generation); 8B=80,469 downloads E3, 3B=38,858 E2, 30B=38,763 E1.
• org=ibm-granite · Granite Speech 5.0 2B multilingual ASR/AST (en/fr/de/es/pt/ja, 390k hrs) P6, plus 470m-turboctc (58,154 downloads) E4; note the -nc variant ships under cc-by-nc-sa-4.0 E8.
• org=ibm-granite · Granite Switch 4.1 8B/30B modular single-endpoint (base + decision + guardian-core adapters + speech encoder), Apache-2.0 P10P11E6E7.
• org=ibm-granite · Granite Timeseries ensemble r1 (PatchTST-FM r1/r2 + FlowState + TTM-r3, GiftEval) P19E17; granite-tsfm v0.3.10/v0.3.11 E23E12.
• org=ibm-granite · deployment/quantization artifacts: bf16-mlx variants E36E37E38 and repeated gguf test tags E43-E57.
• org=ibm-granite · granitelib-core-r1.0 decision LoRA/aLoRA adapters merged for granite-4.1-8b/30b W2.

Talking

• org=IBM · Agentic-AI governance: 80% of EMEA executives saw an AI incident reach board/C-suite in the past year, only 40% changed policy; 75% expect agent autonomy by 2027 P12E14.
• org=IBM · Sovereignty/self-hosted: IBM Bob agentic dev platform gains on-prem/private/sovereign/air-gapped deployment P23E25.
• org=IBM · Open inference: llm-d 1.0, CNCF-governed distributed inference, cited at 5-10x cost reductions on context-heavy agentic workloads P9E13.
• org=IBM · Earnings: Q3 revenue $17.1B (+5%, +6% cc), Software +8%, 'Super Intelligence' client deployments P8E11.
• org=IBM · Workforce/skills: CHRO study (critical thinking/judgment gaps), CFO study (guardrails by 2030), K-12 readiness gap P28P24P3.
• org=IBM · Quantum: DARPA QBI Stage C; HRL acquisition; Anderon $1B CHIPS award; ETH Zurich hub P14E47E31E33.
• org=IBM · Semiconductors: nanostack + Siemens; dual-architecture IBM Z/LinuxONE processor P13E58.
• org=IBM · GTM verticals: SAP ERP, Providence hiring admin (-90%), Swift digital-asset banking, Lightwell open-source remediation P17P15P25P18.
• org=IBM · Open science: NASA-IBM Lunar Foundation Model P2E32; external podcast discussion of agent incident forensics W6.

Shipping

Inspectable artifacts this window: Granite 4.2 3B/8B/30B reasoning models E1E2E3; a three-model Granite Speech 5.0 refresh (2B, revised turboctc, updated 4.1-2b-plus) P6W4E4E5; Granite Switch 4.1 8B/30B P10P11E6E7; Granite Timeseries ensemble r1 plus granite-tsfm v0.3.10/v0.3.11 P19E23E12; and bf16-mlx / gguf deployment artifacts E36E37E38E43-E57. Beyond models, llm-d 1.0 shipped as an open distributed-inference release P9, and IBM Bob added self-hosted deployment P23.

Research themes

• Reasoning/RL: Granite 4.2 8B uses a three-stage pipeline (SFT → multi-environment GRPO over math/code/tool/structured output → preference alignment) with three selectable reasoning modes; 86.67 AIME 2025, 47.67 SWE-Bench Verified W3.
• Modular/multi-capability composition: Switch 4.1 routes chat, decisions, and speech through one endpoint via on-demand adapters P10P11W1W5.
• Speech: multilingual ASR/AST across six languages, robustness to noise/reverberation, keyword-biased ASR up to 2000 keywords P6.
• Time-series: ensemble probabilistic forecasting combining four architectures to reduce single-model failure modes P19.
• Agentic governance/safety: board-level incident escalation and the governance-autonomy gap P12P1; Marist z17 incubator stress-tests agent guardrails under deceptive/unethical tactics P27E29.
• Quantum: fault-tolerant target 2029, DARPA verification, quantum protein simulation (Gordon Bell finalist) P14E35.
• Semiconductors: 3D nanosheet nanostack (~100B transistors at 7Å) with Siemens EDA P13.

Hiring & scaling

Direct hiring evidence is absent from this pack - no role titles, teams, or locations are cited. Scaling signals instead run through M&A and partnerships: Logiq Consulting (UK cybersecurity) P26, HRL Laboratories (quantum) E47, the Anderon quantum-foundry CHIPS award E31, and GTM collaborations with SAP, Providence, and Siemens P17P15P13. Treat the absence of open-role data as a gap in this pack, not evidence of headcount stasis.

Category implications

• Strategy: the sovereignty + governance double-framing (self-hosted Bob, EMEA governance study) positions Granite/watsonx for regulated, data-sensitive, and public-sector buyers P23P12P26.
• Infrastructure: llm-d 1.0 targets inference-cost and throughput pain for agentic workloads, while Bob self-hosting extends control to on-prem/air-gapped estates P9P23; nanostack and the IBM Z dual-architecture processor signal continued compute/device-scale investment P13E58.
• Product: Switch's single-endpoint chat+decisions+speech design reduces serving complexity, and the bf16-mlx/gguf artifacts signal local/edge deployment intent P10P11E36E43-E57.
• Research: active threads are reasoning RL, speech robustness, time-series ensembles, agentic governance/safety, quantum, and semiconductors W3P6P19P12P27P14P13.
• Hiring: no direct signal; infer demand for inference/serving, agentic-safety, speech, and time-series engineering from shipping cadence, but do not over-extrapolate E12E23P6P19.
• GTM: vertical playbook via SAP ERP, healthcare (Providence), sports sponsorships (US Open/Ferrari), digital banking (Swift), and open-source security (Lightwell) P17P15P4P5P25P18.

Traction highlights

• Granite 4.2 8B: 80,469 downloads E3; 3B: 38,858 E2; 30B: 38,763 E1.
• Granite Speech 470m-turboctc: 58,154 downloads E4; -nc variant: 32,797 E8; Speech 5.0 2B (newer): 1,125 E5.
• Granite Timeseries PatchTST-FM-r2: 237,738 downloads E22; swash-3b: 8,916 E41; Switch 4.1 8B/30B: 1,805 / 1,364 E6E7.
• Q3 2026: revenue $17.1B (+5%, +6% cc), Software +8%, FCF outlook raised P8.
• llm-d cited as adopted by 'frontier model labs, hyperscalers and neoclouds, hardware accelerators, and major enterprises' P9.
• Quantum footprint: 340+ organizations across industries P14.