IBM (Granite)

{"content": "## Thesis\n\nIBM is converging its Granite open-source model family with a multi-billion-dollar infrastructure and security play to position as the enterprise-trusted AI platform. The lab ships Apache 2.0-licensed models across seven modalities — language, code, vision, speech, time series, embeddings, and geospatial — while simultaneously building the hardware substrate (sub-1nm chips) and supply-chain security infrastructure (Project Lightwell, $5B+ commitments) that enterprises need to deploy AI with sovereignty and control P2P3E13P17. The Granite Switch pattern — embedding 12 task-specialized LoRA adapters into a single checkpoint — signals a deployment-efficiency thesis: one model, multiple capabilities, no model-switching overhead W1W2. Partner integrations with ServiceNow, Google Cloud, OpenAI, Deloitte, and Palo Alto Networks extend distribution and create GTM pathways for consulting-led AI adoption P10E31P5P2P4. IBM's own research reinforces the urgency: 91% of executives report not fully understanding their AI dependencies, and 71% say switching AI vendors would be difficult — precisely the lock-in problem IBM's open-source, multi-model strategy is designed to address P8E25.\n\n## Signal desks\n\n### Hiring\n\nNo cited evidence in this pack. The evidence set contains no job postings, role descriptions, team growth announcements, or hiring-related signals for IBM's Granite or AI organizations.\n\n### Forks\n\nNo cited evidence in this pack. All documented repositories in the ibm-granite GitHub org are first-party creations. No fork events from IBM to upstream projects (e.g., PyTorch, vLLM, transformers, agent frameworks, eval harnesses) appear in the evidence. The one external fork observed (rattnak/granite-switch) is a community fork *of* IBM, not by IBM W3.\n\n### Releases\n\n- Granite 4.1 language model family (April 6, 2026): 3B, 8B, and 30B dense models shipped with base and instruct variants under Apache 2.0; the 8B instruct variant reached 432K HuggingFace downloads E2E4E7E34E35E38.\n- Granite Switch 4.1 previews (May 1, 2026): 3B, 8B, and 30B checkpoints each embedding 12 task-specialized LoRA adapters covering Core (requirement check, context attribution, uncertainty), RAG (query rewrite, clarification, answerability, hallucination detection, citation), and Guardian (safety, factuality, policy) libraries E14E16E22W2.\n- Granite Speech 4.1 (Feb–April 2026): 1B and 2B automatic speech recognition models released; the 2B variant reached 418K downloads and the non-autoregressive (NAR) variant reached 178K downloads E1E3E9E10.\n- Granite Vision 4.1-4B (April 16, 2026): LoRA adapter on Granite-4.1-3B for enterprise document data extraction — chart-to-CSV, table extraction, KVP extraction, image-to-text; accompanied by the ChartNet dataset (1.7M synthetic charts) E8P25.\n- Granite Guardian 4.1-8B (April 16, 2026): Safety/risk detection model adding improved Bring Your Own Criteria (BYOC) support for arbitrary judging criteria beyond pre-baked safety detectors; predecessor Granite Guardian 3.3 ranked #3 on LLM-AggreFact and #1 on REVEAL E21P19.\n- Time Series: granite-tsfm v0.3.6 shipped May 2026; Granite-Timeseries-PatchTST released March 2026 with 21K downloads E41E49.\n- Specialized releases: Granite-3.3-8B-math-prm-v2 (process reward model for math) E40; Granite-3.3-8B-security-lib E47; granitelib-rag-gpt-oss-r1.0 E43; Granite Guardian Japanese toxicity variant E46; factuality detection and correction LoRA adapters E50E52.\n- New repos created: granite.build (LLM pipeline orchestration, Python, 41 stars) E42; granite.debug-tools (8 stars) E45; granite.trust.policy-tools (Jupyter Notebook, 16 stars) E48; granite-speech (empty template, July 1, 2026 — may presage a new speech initiative) P1E15; granite-4.1-language-models (87 stars) E44.\n- Deprecation: granite-io archived May 2026; functionality folded into Mellea (generative-computing/mellea) P22E51.\n\n### Talking\n\n- Project Lightwell / supply chain security (dominant theme, June 2026): IBM, Red Hat, and Deloitte announced a collaboration to scale automated vulnerability patching across regulated software supply chains, decoupling security remediation from upgrade cycles P2E17. Two days earlier, Palo Alto Networks joined to integrate virtual patching with Lightwell's software remediation P4E19. A May 28 announcement framed the $5B IBM-Red Hat commitment around Lightwell as \"a trusted enterprise clearinghouse for open source software\" E13.\n- OpenAI Daybreak Cyber Partner Program (June 22, 2026): IBM joined OpenAI's program to deploy frontier AI models defensively in enterprise security workflows; launched a managed application security service using OpenAI cyber capabilities with read-only code access and bounded execution via IBM Consulting Advantage P5E20.\n- Sub-1 nanometer chip breakthrough (June 25, 2026): IBM unveiled 0.7nm (7 angstrom) chip technology with \"nanostack\" 3D architecture, packing nearly 100B transistors — ~2x density of IBM's 2nm node, projected 50% more performance or 70% greater energy efficiency P3E18.\n- AI sovereignty and control gap: IBM Institute for Business Value study of 1,000 senior executives reported 68% struggle with data residency/sovereignty requirements, 91% don't fully understand AI dependencies, 71% say switching AI vendors/models would be difficult, and organizations with advanced AI control capabilities protect more than half of operating profit from AI-driven disruptions P8E25. A companion study focused on CIOs/CTOs facing a growing \"AI control gap\" E29.\n- IBM Z critical infrastructure (June 19, 2026): Announced GA of three new Z software tools — zSecure Detection, and others — designed to address frontier model attacks and complement Project Glasswing and Lightwell P7E24.\n- Enterprise partnerships: Multi-year ServiceNow collaboration to modernize legacy systems and unlock enterprise data for agentic AI via ServiceNow Workflow Data Fabric combined with IBM data capabilities P10E27. Google Cloud strategic partnership to scale AI with IBM Consulting Advantage and industry-specific agents for Gemini Enterprise E31.\n- Wimbledon 2026 AI experiences: New Key Moments tool (explaining match momentum with AI reasoning) and Match Chat (natural-language interactive companion) on watsonx Orchestrate, with fully redesigned digital platforms P6E23.\n- Apptio AI financial operations: Conversational Insights preview enabling plain-language queries about technology spend across hybrid IT estates; new capabilities for Data Center TCO, Cloudability Intelligent Forecasting, and Advanced Containers P9E26.\n- Quantum computing: $10B+ five-year commitment spanning R&D, manufacturing, M&A, and ecosystem expansion toward fault-tolerant quantum computers E33.\n- Education push: IBM Bob global AI Builders Challenge expanded to 20,000 post-secondary institutions worldwide E32.\n- Granite Switch narrative: IBM Research blog framed Switch as making \"unpredictable text generation into a reliable, deterministic programming function\" via the Mellea open-source library for generative computing W1. External analysis characterized the 12-adapter-in-one-checkpoint pattern as \"the deployment story\" W2.\n\n## Shipping\n\nThe Granite 4.1 family (3B/8B/30B dense, plus base variants) shipped on HuggingFace April 6, 2026, with the 8B instruct model accumulating 432K downloads E2E4E7. The Granite Switch 4.1 previews followed on May 1, 2026, packing 12 LoRA adapters into single checkpoints across all three scales E14E16E22W2. Modality expansions shipped concurrently: Granite Speech 4.1 models (1B and 2B ASR, including NAR and plus variants) reached 418K downloads on the 2B model E3E1E9E10; Granite Vision 4.1-4B shipped as a LoRA adapter for enterprise document extraction alongside the ChartNet dataset (1.7M synthetic charts, 94K human-verified) E8P25; Granite Guardian 4.1-8B shipped with improved BYOC support E21P19. Infrastructure tooling shipped via new repos: granite.build for LLM pipeline orchestration E42, granite.debug-tools E45, and granite.trust.policy-tools E48. The granite-io library was deprecated and rolled into Mellea P22. A new granite-speech repo appeared July 1, 2026 as an empty template, suggesting an upcoming speech initiative beyond the existing granite-speech-models repository P1E15. The granite-tsfm library shipped v0.3.6 E41.\n\n## Research themes\n\nMulti-modal foundation models as enterprise building blocks. IBM's research posture treats modalities as independent foundation model problems: language (Granite 3.0→3.1→3.3→4.0→4.1 lineage) P17P23P26E2, code (116-language code intelligence models) P12, vision (document extraction via LoRA adapters on language backbones) P25, speech (two-pass ASR/AST design) P28, time series (PatchTSMixer, TinyTimeMixer, PatchTST foundation models) P11, embeddings (bi-encoder R2 models with retrieval-oriented pretraining and contrastive finetuning) P24, and geospatial (biomass, land surface temperature, canopy height) P16P18P20P21.\n\nSupply-chain security as a first-class research problem. Project Lightwell frames open-source vulnerability remediation as a research and engineering challenge requiring coordinated upstream threat disclosure, independent maintainer coordination, and backported patches to pinned production versions P2P4. The OpenAI Daybreak Cyber partnership adds frontier-AI-driven vulnerability identification and validation to this pipeline P5E20.\n\nDeterministic generative computing. The Granite Switch + Mellea architecture reframes LLM calling as a programming paradigm: control tokens activate task-specialized LoRA adapters per-token, turning stochastic generation into deterministic, auditable function calls W1W2. This theme also appears in the Granite 3.3 language models' introduction of fill-in-the-middle (FIM) support and reasoning-trace separation (intermediate thoughts vs. final answers) P26.\n\nSafety and factuality as a model family, not a filter layer. Granite Guardian has evolved across four major revisions (3.0→3.2→3.3→4.1), adding verbalized confidence, hybrid thinking mode, multi-risk detection, harm-correction LoRA adapters, factuality detection/correction, and BYOC support for arbitrary judging criteria P19E21E50E52. IBM publicly benchmarks these models, noting Guardian 3.3 8B outperforms GPT-4o and Mistral Large 2 on LLM-AggreFact despite its smaller size P19.\n\nHardware-software co-design for the AI era. The 0.7nm nanostack chip architecture P3E18 and IBM Z software tools purpose-built for \"frontier model attacks\" P7E24 signal a research agenda that treats AI infrastructure as a full-stack problem from transistor to mainframe.\n\nSynthetic data as a training strategy. Granite 3.3 language models were \"trained on synthetic data generated from a variety of different open source LLMs, including but not limited to open source models like Mistral and Gemma\" P26. ChartNet is a million-scale synthetic dataset for chart understanding P25. Granite Speech models use \"synthetic datasets tailored to support the speech translation task\" P28.\n\n## Hiring & scaling\n\nNo cited evidence in this pack. The evidence contains no job postings, headcount disclosures, team structure descriptions, location expansion signals, or hiring announcements for IBM's AI, Granite, or research organizations. IBM's public announcements focus on product releases, partnerships, and capital commitments ($5B for open-source AI with Red Hat E13, $10B+ for quantum E33) but do not disclose the associated hiring plans or team scaling trajectories in this evidence set. This is a notable gap for a lab shipping models across seven modalities while simultaneously building new infrastructure (Lightwell) and hardware (sub-1nm chips, IBM Z tools).\n\n## Category implications\n\nInfrastructure. The 0.7nm chip P3E18 and IBM Z software for frontier model attacks P7E24 position IBM to sell AI compute and security infrastructure directly, not just models. The nanostack 3D architecture claim — 50% more performance or 70% greater energy efficiency over the prior node — creates a hardware differentiation story that no other AI lab in this evidence set can match P3. Apptio's new Data Center TCO and Cloudability tools extend this into the financial governance layer for AI infrastructure spend P9E26.\n\nProduct. The Granite product strategy is breadth-over-giant-scale: seven modalities, Apache 2.0 licensing throughout, and models small enough to run on-device or on constrained compute P17P12P25P28P11P24. The Granite Switch pattern — single checkpoint, 12 adapters, per-token routing — is a deployment-product innovation that reduces the operational complexity of managing multiple specialized models W1W2. Watsonx Code Assistant Individual brings Granite code models to local VS Code via Ollama, signaling a developer-tool product vector P14.\n\nResearch. IBM's research posture is applied-enterprise: each modality targets a specific enterprise use case (code for development, vision for document extraction, speech for ASR/AST, time series for forecasting, geospatial for earth observation) rather than pursuing AGI. The synthetic-data training strategy P26P25P28 and the Guardian safety-model-as-product approach P19 suggest research is organized around enterprise deployability, not benchmark-chasing. The ChartNet dataset release (1.7M synthetic charts with aligned code, tables, summaries, and reasoning) is a research artifact that also serves as a product differentiator for the vision models P25.\n\nStrategy. Project Lightwell is a strategic wedge: by solving open-source supply-chain security at scale, IBM creates an enterprise dependency that extends beyond model adoption into infrastructure and services P2P4E13. The partnership cascade — Red Hat (platform), Deloitte (services integration), Palo Alto Networks (network security), OpenAI (frontier AI for cyber), ServiceNow (workflow automation), Google Cloud (cloud distribution) — builds a consulting-and-integration moat around Granite that pure model labs cannot replicate P2P4P5P10E31. IBM's own survey research creates the demand narrative: 91% of executives lack AI dependency visibility, 71% are locked into vendors they can't easily switch P8E25. The strategy is to be the open, auditable, sovereign alternative.\n\nHiring. The evidence gap here is itself a signal: IBM is making massive capital commitments ($5B open-source AI, $10B+ quantum) E13E33 while shipping models across seven modalities, yet no hiring plan is publicly disclosed. This may imply reliance on existing IBM Research and Red Hat teams rather than a net-new hiring wave, or that hiring is handled through non-public channels. Either interpretation warrants monitoring.\n\nGTM. IBM's GTM motion is enterprise consulting-led: the Wimbledon partnership showcases consumer-facing AI P6E23, but the core distribution is through IBM Consulting Advantage (the AI-powered delivery platform), ServiceNow workflows P10, Google Cloud's Vertex AI E31, and Apptio's financial governance tools P9. The IBM Bob education initiative seeding Granite at 20,000 institutions E32 is a long-term developer-adoption play. Abertis mobility modernization E37 demonstrates the classic IBM GTM pattern: five-year, multi-subsidiary digital transformation deals that embed IBM technology into core operations.\n\n## Traction highlights\n\n- Granite-4.1-8B (instruct): 432,773 HuggingFace downloads, 203 likes E2\n- Granite-speech-4.1-2B: 418,410 HuggingFace downloads, 146 likes E3\n- Granite-4.1-3B (instruct): 336,140 HuggingFace downloads, 85 likes E7\n- Granite-speech-4.1-2B-nar: 178,004 HuggingFace downloads E10\n- Granite-4.0-1B-speech: 111,235 HuggingFace downloads, 249 likes E1\n- Granite-code-models repo: 1,248 GitHub stars, 82 forks P12\n- Granite-tsfm repo: 860 GitHub stars, 275 forks P11\n- Granite-3.0-language-models repo: 270 GitHub stars P17E53\n- Granite-4.0-language-models repo: 212 GitHub stars E54\n- Granite-guardian repo: 152 GitHub stars P19E55\n- Granite-3.1-language-models repo: 145 GitHub stars P23E56\n- Granite-vision-4.1-4B: 86,857 HuggingFace downloads, 92 likes E8\n- Granite-4.1-language-models repo (new, April 2026): 87 GitHub stars E44\n- Granite-embedding-models repo: 71 GitHub stars P24E57\n- Granite-guardian-4.1-8B: 51,449 HuggingFace downloads E21\n- Granite-vision-4.0-3B: 54,856 HuggingFace downloads E5\n- Granite.build repo (new, April 2026): 41 GitHub stars E42\n- Watsonx-code-assistant-individual repo: 37 GitHub stars P14\n- Project Lightwell: $5B IBM-Red Hat commitment E13; Deloitte and Palo Alto Networks as named integration partners P2P4\n- OpenAI Daybreak Cyber: IBM named as partner; new managed application security service launched P5E20\n- Hacker News traction on $5B open-source announcement: 5 points E13\n- Granite Guardian 3.3 benchmark standing: #3 on LLM-AggreFact, #1 on REVEAL, outperforming GPT-4o and Mistral Large 2 despite 8B parameters P19\n\n## Sources\n\nP1 ibm-granite/granite-speech repo (July 2026)\nP2 IBM, Red Hat, and Deloitte Announce Lightwell Collaboration (June 26, 2026)\nP3 IBM Debuts World's First Sub-1 Nanometer Chip Technology (June 25, 2026)\nP4 IBM, Red Hat and Palo Alto Networks Expand Project Lightwell (June 24, 2026)\nP5 IBM and OpenAI Bring Frontier AI to Cyber Defense (June 22, 2026)\nP6 Wimbledon and IBM Introduce New AI-Powered Fan Experiences (June 22, 2026)\nP7 Protecting and Innovating Critical Infrastructure Through New Security Landscapes (June 19, 2026)\nP8 IBM Study: Limited Control and Rising Dependencies Leave Enterprises Exposed (June 17, 2026)\nP9 Apptio Unveils Conversational Insights (June 16, 2026)\nP10 IBM and ServiceNow Expand Collaboration (June 11, 2026)\nP11 ibm-granite/granite-tsfm repo (860 stars)\nP12 ibm-granite/granite-code-models repo (1,248 stars)\nP13 ibm-granite/discussions repo\nP14 ibm-granite/watsonx-code-assistant-individual repo\nP15 ibm-granite/.github repo\nP16 ibm-granite/granite-geospatial-biomass repo\nP17 ibm-granite/granite-3.0-language-models repo (270 stars)\nP18 ibm-granite/granite-geospatial-land-surface-temperature repo\nP19 ibm-granite/granite-guardian repo (152 stars)\nP20 ibm-granite/granite-geospatial-canopyheight repo\nP21 ibm-granite/geospatial repo\nP22 ibm-granite/granite-io repo (deprecated, archived)\nP23 ibm-granite/granite-3.1-language-models repo (145 stars)\nP24 ibm-granite/granite-embedding-models repo (71 stars)\nP25 ibm-granite/granite-vision-models repo (46 stars)\nP26 ibm-granite/granite-3.3-language-models repo (23 stars)\nP27 ibm-granite/granite-common repo\nP28 ibm-granite/granite-speech-models repo (44 stars)\nE1 granite-4.0-1b-speech release (111K downloads)\nE2 granite-4.1-8b release (432K downloads)\nE3 granite-speech-4.1-2b release (418K downloads)\nE4 granite-4.1-30b release (77K downloads)\nE5 granite-4.0-3b-vision release (54K downloads)\nE6 granite-4.0-h-350m release\nE7 granite-4.1-3b release (336K downloads)\nE8 granite-vision-4.1-4b release (86K downloads)\nE9 granite-speech-4.1-2b-plus release\nE10 granite-speech-4.1-2b-nar release (178K downloads)\nE11 granite-4.0-350m release\nE12 granite-4.0-1b release\nE13 IBM and Red Hat Commit $5 Billion to Open Source in AI Era (May 28, 2026)\nE14 granite-switch-4.1-3b-preview release\nE15 granite-speech repo creation (July 1, 2026)\nE16 granite-switch-4.1-8b-preview release\nE17 Deloitte Lightwell collaboration announcement (June 26, 2026)\nE18 Sub-1nm chip announcement (June 25, 2026)\nE19 Palo Alto Networks Lightwell expansion (June 24, 2026)\nE20 OpenAI Daybreak Cyber announcement (June 22, 2026)\nE21 granite-guardian-4.1-8b release (51K downloads)\nE22 granite-switch-4.1-30b-preview release\nE23 Wimbledon AI announcement (June 22, 2026)\nE24 IBM Z critical infrastructure post (June 19, 2026)\nE25 AI sovereignty study (June 17, 2026)\nE26 Apptio Conversational Insights (June 16, 2026)\nE27 ServiceNow collaboration (June 11, 2026)\nE28 granite-4.0-h-350m-base release\nE29 CIO/CTO AI control gap study (June 8, 2026)\nE30 granite-4.0-1b-base release\nE31 Google Cloud partnership (June 4, 2026)\nE32 IBM Bob AI Builders Challenge (June 3, 2026)\nE33 $10B Quantum Computing commitment (June 2, 2026)\nE34 granite-4.1-30b-base release\nE35 granite-4.1-8b-base release\nE36 granite-4.0-350m-base release\nE37 Abertis mobility modernization (May 26, 2026)\nE38 granite-4.1-3b-base release\nE39 granite-vision-3.3-2b-chart2csv-preview release\nE40 granite-3.3-8b-math-prm-v2 release\nE41 granite-tsfm v0.3.6 release\nE42 granite.build repo (41 stars)\nE43 granitelib-rag-gpt-oss-r1.0 release\nE44 granite-4.1-language-models repo (87 stars)\nE45 granite.debug-tools repo (8 stars)\nE46 granite-guardian-4.0-3b-toxicity-ja release\nE47 granite-3.3-8b-security-lib release\nE48 granite.trust.policy-tools repo (16 stars)\nE49 granite-timeseries-patchtst-fm-r1 release (21K downloads)\nE50 granite-guardian-3.2-8b-factuality-detection release\nE51 granite-io v0.5.3 release\nE52 granite-guardian-3.2-5b-lora-factuality-correction release\nE53 granite-3.0-language-models repo (270 stars)\nE54 granite-4.0-language-models repo (212 stars)\nE55 granite-guardian repo (159 stars at creation)\nE56 granite-3.1-language-models repo (147 stars)\nE57 granite-embedding-models repo (75 stars)\nE58 granite-io repo (56 stars)\nE59 granite-speech-models repo (45 stars)\nE60 granite-vision-models repo (45 stars)\nW1 IBM Research blog: How to build AI more like software (June 4, 2026)\nW2 The AI Bench: Granite-Switch 4.1 fast take (May 25, 2026)\nW3 rattnak/granite-switch community fork (May 29, 2026)\nW4 MC Press Online: Granite 3.0 coverage\nW5 IBM Granite GitHub org overview", "cites": ["P1", "P2", "P3", "P4", "P5", "P6", "P7", "P8", "P9", "P10", "P11", "P12", "P14", "P16", "P17", "P18", "P19", "P20", "P21", "P22", "P23", "P24", "P25", "P26", "P28", "E1", "E2", "E3", "E4", "E5", "E7", "E8", "E9", "E10", "E13", "E14", "E15", "E16", "E17", "E18", "E19", "E20", "E21", "E22", "E23", "E24", "E25", "E26", "E27", "E29", "E31", "E32", "E33", "E34", "E35", "E37", "E38", "E40", "E41", "E42", "E43", "E44", "E45", "E46", "E47", "E48", "E49", "E50", "E51", "E52", "E53", "E54", "E55", "E56", "E57", "E58", "E59", "E60", "W1", "W2", "W3"]}</｜｜DSML｜｜parameter> </｜｜DSML｜｜invoke> </｜｜DSML｜｜tool_calls>