IBM (Granite)

Thesis

IBM's Granite program is running two distinct strategies in parallel. On the models side, the lab is pushing a broad open-weights portfolio — text, speech, vision, time series, and safety models — mostly Apache 2.0, now shifting from instruction-following assistants to dense reasoning models with native thinking modes and agentic reinforcement learning W1W2W4W6E2E3E5. On the corporate side, energy and capital are flowing into quantum (the HRL acquisition, a Swiss quantum hub at ETH Zurich, a Gordon Bell finalist run, and a modular cryogenics milestone) and into enterprise distribution through an OpenAI consulting partnership and a Red Hat/LTM remediation collaboration P28E25P5P7E47E60P6. The net read: IBM is positioning Granite as the open-weights layer for enterprise agents while betting that durable differentiation lives in quantum hardware and trusted, regulated deployment rather than frontier-language scale.

Signal desks

Hiring — IBM Research is recruiting scientists and engineers into a “Core Foundation Models” team tied to the open-source Granite project, with role language pointing to agentic AI models and reinforcement-learning platforms W5. Evidence is a single job-aggregation source with no location or headcount detail, so this is a thin leading signal W5.

Forks — No cited evidence in this pack.

Releases — Granite 4.2 language models (3B/8B/30B) shipped under Apache 2.0 with thinking/non-thinking and low-effort modes E2E3E5W1W2. Deployment variants followed quickly: MLX for Apple Silicon and fp8/nvfp4/mxfp4 quantizations P14P15P16E49E51E52E58E59. The ibm-granite/gguf repo shows active GGUF/Ollama packaging, including a “test-v5.0-oss-120B” series implying a 120B open-source model under test E30E31E46E44. Adjacent family releases: speech 4.1 and 5.0 turboctc, vision 4.1-4b, guardian 4.1-8b, switch 4.1-3b-preview, and time series PatchTST-FM-r2 E1E6E7E4E13E40E19. Tooling: granite-common gained RAG intrinsics, gpt-oss support, and aLoRA/LoRA adapter handling P13P17P18P19.

Talking — IBM's public writing concentrates on four themes: (1) scientific AI, via the NASA-IBM Lunar Foundation Model P4E9; (2) quantum, via the HRL acquisition, the ETH Zurich hub, the Gordon Bell finalist, and cryogenic milestones P28E25P5P7E47; (3) enterprise security and GTM, via the OpenAI partnership and Lightwell E60P6E11; and (4) applied brand AI, via the US Open fan experience and the K-12 readiness study/fellowship P23P3P20E14.

Shipping

The shipping cadence splits into a headline model family and a long tail of deployment formats. Granite 4.2 is the anchor — three dense decoder-only sizes with reasoning, released under Apache 2.0 W1W2E2E3E5. IBM immediately proliferated formats: MLX for Apple Silicon and fp8/nvfp4/mxfp4 quantizations, signaling an edge/local-inference distribution push P14E49E51E52E58E59. A parallel GGUF pipeline is iterating through Ollama library release candidates and language/speech/120B test tags E46E44E30E21. On the non-language side, speech and vision are the volume shippers (see Traction), and time series got a SOTA-claiming release with PatchTST-FM-r2 plus a granite-tsfm v0.3.9 tooling update E1E4P21P22. The scientific/quantum track ships through research collaborations rather than weights — the lunar foundation model and the quantum supercomputing result — so its release artifacts are reports and PR, not model cards P1P4P7.

Research themes

• Reasoning and agentic RL — Granite 4.2 is IBM's first dense reasoning family; the 8B and 30B are post-trained via RL inside code, terminal, and web-search environments W2W4W6.
• Time-series foundation models — PatchTST-FM-r2 introduces Conformer blocks, overlapping patches with Hamming-window loss weighting, and CauKer-generated synthetic data; it ranks #1 on GIFT-Eval among replicable zero-shot models P21W3.
• Scientific foundation models — the NASA-IBM Lunar Foundation Model turns multi-instrument lunar observation data into a general scientific base model P4P1.
• Quantum-centric supercomputing — a 12,635-atom protein simulation via quantum-plus-classical (Heron processors with Fugaku/Miyabi-G) P7; silicon-spin qubits added via HRL P28.
• Multimodal speech/vision — turboctc ASR and vision 4.1 extend Granite beyond text E7E4.
• Safety and GRC — Granite Guardian 8B, plus governance/risk/compliance evaluation language baked into Granite 4.2 model cards E13P14.

Hiring & scaling

Thin evidence: the only hiring signal is an IBM Research “Core Foundation Models” team building the Granite project, with positions framed around agentic AI models and RL platforms W5. There is no cited evidence on locations, team size, eval infrastructure, or data/commercialization roles — the pack proves the gap rather than filling it W5.

Category implications

• Strategy — IBM is running quantum and open-weights AI as complementary bets: HRL adds silicon-spin qubits to the superconducting roadmap (Starling 2029, Blue Jay mid-2030s) P28, while Granite 4.2 targets the enterprise-agent market W1W6.
• Infrastructure — quantum System Two/Nighthawk at CSCS and modular cryogenic systems signal hardware-led differentiation P5E47; dual-architecture IBM Z/LinuxONE broadens Arm-native enterprise compute E36; fp8/nvfp4/mxfp4 and MLX variants push local/edge inference E49E51P14.
• Product — the Granite suite is consolidating text + speech + vision + time series + guardrails, with GGUF/Ollama and Confluent Early Access as distribution channels E13E46W3.
• Research — IBM is front-running two applied frontiers — SOTA zero-shot time series and scientific foundation models — where its enterprise/data advantage shows P21P4.
• Hiring — demand is clustering on agentic AI and RL engineering inside the foundation-models org W5.
• GTM — an OpenAI partnership embeds GPT-5.6 into IBM Consulting, while Red Hat/LTM Lightwell and the US Open/education programs build enterprise-security and brand reach E60P6P23P20.

Traction highlights

• Speech is the volume leader: granite-speech-4.1-2b at ~362k downloads and 4.1-2b-plus at ~125k E1E6.
• Vision 4.1-4b at ~116k downloads E4; Guardian 4.1-8b at ~31k E13.
• Granite 4.2 text: 3B ~26.6k, 8B ~26.8k, 30B ~15k downloads E3E5E2.
• Time series PatchTST-FM-r2 at ~2.6k downloads with a #1 GIFT-Eval rank E19P21.
• Quantized/MLX variants are long-tail (≤~120 downloads), consistent with an early-stage edge deployment push E15E16E17E49.