IBM (Granite) analysis

Thesis

IBM is executing two parallel strategies in this pack. First, it is shipping an openly licensed Granite family spanning text, speech, and time-series models — 3B/8B/30B LLMs with native reasoning, a 470M TurboCTC ASR model, and forecasting/ensemble releases — framed explicitly for enterprise agentic workflows W1 E1 E4 W3. Second, it is investing heavily in sovereign and regulated deployment (self-hosted IBM Bob, on-prem Digital Asset Haven, Lightwell open-source remediation, Logiq cybersecurity) and in quantum (DARPA QBI Stage C, a $1B CHIPS-funded foundry, the HRL acquisition, a Gordon Bell finalist) P13 P15 P8 P16 P4 P19 E37 P23. Hiring and fork evidence is thin in this pack, so shipping cadence and capital allocation are the strongest leading indicators rather than open requisitions W5.

Signal desks

Hiring — IBM Research is advertising Research Scientist and Engineer roles on its Core Foundation Models team working on Granite, spanning agentic AI models and RL platforms; the listing is third-party and discloses no location or headcount W5. No other role-level hiring signals are cited in this pack.

Forks — No cited evidence in this pack. The visible GitHub activity (granite-tsfm, granite-common, gguf, granite-4.2-language-models) is first-party release and repo activity rather than forks of upstream projects P10 P24 E51.

Releases — Granite 4.2 LLMs shipped in 3B/8B/30B (Apache-2.0) with native reasoning for agents E1 E2 E3 W1. Granite Speech 5.0 TurboCTC 470M ASR (Apache-2.0) shipped, plus a cc-by-nc-sa non-commercial variant E4 E6 W4. In time series, PatchTST-FM-r2 (~385M, OpenMDW-1.0) and the Granite Timeseries Ensemble R1 framework (Apache-2.0) combining four checkpoints shipped E7 W3 P9 E17. Tooling releases include granite-tsfm v0.3.10 (adds the ensemble) P10 E13, granite-common packaging up to v0.3 P24 P28, and repeated gguf quantization/test tags for Ollama, MLX, FP8, and NVFP4 variants E28 E29 E30 E33 E56 E60.

Talking — Research writing frames Granite 4.2 as native reasoning for enterprise agents W1, and an IBM Technology panel ties it to the frontier-pacing debate W6. Speech and time-series posts emphasize permissive licensing, the speed-accuracy Pareto frontier, and commercial-friendly licenses W2 W4 W3. The corporate narrative is dominated by quantum (DARPA Stage C, $1B CHIPS award, HRL acquisition, Gordon Bell finalist, ETH Zurich hub) P4 P19 E37 P23 P21, sovereignty and enterprise security (Bob self-hosted, Lightwell plus Clearinghouse GA, Logiq acquisition, Digital Asset Haven on-prem) P13 P8 P16 P15, and market-readiness surveys targeting CFOs, CHROs, K-12, and sports fans P14 P18 P2 P3. Q3 2026 earnings are set for Oct 21 P6 E10.

Shipping

Concrete artifacts in this pack:

Granite 4.2 LLMs on Hugging Face (Apache-2.0): 30B at 38,753 downloads / 127 likes, 3B at 49,921 / 110, 8B at 132,068 / 97 E1 E2 E3.
Quantized deployment variants: bf16-MLX (3B/8B/30B), FP8 (3B), NVFP4 (8B) E28 E29 E30 E56 E60.
Granite Speech 5.0 TurboCTC (57,988 downloads) plus a TurboCTC-NC variant E4 E6.
Granite Time Series PatchTST-FM-r2 (304,769 downloads) E7 and the ensemble R1 framework (44 downloads, 2 likes) E17.
Repos and releases: granite-tsfm v0.3.10 P10 E13, granite-common v0.1.0 through v0.3 P27 P28, granite-4.2-language-models repo (35 stars) E51, and a series of gguf test tags E33.
Enterprise and other products: IBM Bob self-hosted deployment P13 E15, Lightwell Clearinghouse GA after 400+ remediated Java library vulnerabilities P8 E5, IBM Digital Asset Haven Swift beta and on-prem beta P15 E19, and the open-source NASA-IBM Lunar Foundation Model P20 P1.

Research themes

Agentic reasoning: Granite 4.2 adds native step-by-step reasoning so models plan before acting, with mid-training innovations and synthetic code training via Code Alchemy W1 W6.
Efficient ASR: TurboCTC uses only public data, a Muon optimizer, and inference optimizations to reach the Open ASR speed-accuracy Pareto frontier W2 W4.
Time-series foundation models: probabilistic forecasting, imputation, and ensemble aggregation across PatchTST/FlowState/TTM families, distributed as recipes rather than merged weights P9 W3.
Quantum-centric supercomputing: a 2029 fault-tolerant target, modular cryogenic systems linking hundreds of chips, and a 12,635-atom protein simulation combining quantum and classical methods P4 E58 P23 E26.
Scientific foundation models: the NASA-IBM Lunar Foundation Model for multi-instrument lunar data P20.
Open-source security: Lightwell's AI-driven discovery and backported remediation of 400+ previously unknown Java library vulnerabilities P8 E5.

Hiring & scaling

The only cited open-role signal is IBM Research Foundation Models hiring for Granite (agentic AI, RL platforms); the pack is thin on specific teams, locations, eval/data roles, and GTM hiring W5.
Scaling is instead visible through capital and facilities: the $1B CHIPS award plus $1B IBM investment in the Anderon quantum foundry P19 E22, completion of the HRL Laboratories acquisition E37, the ETH Zurich Quantum System Two hub P21, the Logiq Consulting acquisition for UK cybersecurity P16, and the Marist z17 innovation incubator P17.
Because direct hiring evidence is sparse, treat the quantum and sovereignty acquisitions and facilities as the stronger near-term scaling indicators P19 E37 P16 P21.

Category implications

Strategy: IBM pairs open Granite adoption with sovereignty and regulated-deployment offerings (self-hosted Bob, on-prem Digital Asset Haven, Lightwell, Logiq), an open-at-the-model-layer, controlled-at-the-deployment-layer posture W1 P13 P15 P8 P16.
Infrastructure: outsized quantum hardware investment (wafer foundry, cryogenic fridges, Quantum System Two/Nighthawk) implies a multi-year buildout beyond LLM compute P19 E58 P21; self-hosted and air-gapped AI deployment targets regulated enterprises P13.
Product: the Granite portfolio spans text (3B-30B), speech (470M ASR), and time series (PatchTST/FlowState/TTM), positioning it as an enterprise model catalog rather than a single flagship W1 W4 W3.
Research: effort concentrates on agentic reasoning, efficient and permissive ASR, time-series ensembles, quantum-centric supercomputing, and scientific foundation models W1 W2 P9 P23 P20.
Hiring: thin evidence points to continued Granite FM team buildout but lacks cited eval/data/GTM role detail; monitor future requisitions for those gaps W5.
GTM: consulting-led co-delivery and enterprise partnerships (Providence, SAP, Swift/digital assets, LTM and Confluent Cloud distribution) are the go-to-market motion P5 P7 P15 P22 W3; survey content targets CFOs, CHROs, education, and sports to prime AI adoption narratives P14 P18 P2 P3.

Traction highlights

Granite 4.2 8B leads HF downloads at 132,068 (97 likes); 3B at 49,921; 30B at 38,753 (127 likes) E3 E2 E1.
PatchTST-FM-r2 is the strongest time-series signal at 304,769 downloads E7.
Granite Speech 5.0 TurboCTC at 57,988 downloads E4.
Low-traction caveats: ensemble R1 at 44 downloads / 2 likes E17; the ensemble-forecasting model card returned 'Entry not found' P11; the NC, FP8, and NVFP4 variants show minimal downloads E6 E60 E56.
Validation milestones: 400+ vulnerabilities remediated P8 E5; DARPA QBI Stage C P4 E8; ACM Gordon Bell finalist P23 E26.