On July 15, 2026, Thinking Machines Lab released Inkling under Apache 2.0 — a 975B-parameter Mixture-of-Experts model with 41B active parameters, available in full on Hugging Face and deployable today across Databricks, Baseten, Modal, Fireworks, and TogetherAI.
The open-weights frontier in 2026 has moved faster than most enterprise procurement cycles. Kimi K2.6, DeepSeek V4 Pro, GLM 5.2, and Nemotron 3 Ultra have compressed the gap between proprietary frontier APIs and self-hostable alternatives. Thinking Machines Lab enters this race with a deliberate counter-positioning: Inkling is designed to be a broad, balanced foundation model built for customization, with multimodal breadth as its primary differentiator rather than benchmark supremacy on any single leaderboard. That choice carries precise commercial implications for every enterprise currently locked into a per-token API contract.
Architecture and capabilities
Inkling runs 256 routed experts plus 2 shared experts per MoE layer, with 6 experts active per token. The context window reaches 1 million tokens, trained on 45 trillion tokens of text, images, audio, and video on NVIDIA GB300 NVL72 systems. A sigmoid-based router with auxiliary-loss-free load balancing handles expert selection — a design choice that improves inference stability at scale and lowers per-query latency variance compared to loss-based routing.
Multimodal breadth is Inkling's clearest differentiator in the open-weights category. Audio arrives natively via dMel spectrograms; images encode as 40×40 pixel patches through a four-layer hMLP. The model reasons across all three modalities in a single forward pass. On audio benchmarks, Inkling scores 91.4% on VoiceBench and 77.2% on MMAU. On vision tasks, it reaches 73.5% on MMMU Pro and 78.1% on CharXiv RQ. These are first-release scores on a model explicitly designed for fine-tuning — enterprise-adapted versions will improve further.
Coding and agentic performance rounds out the picture. Inkling posts 77.6% on SWE-Bench Verified and 63.8% on Terminal Bench 2.1 — the latter achieved at one-third the token count of Nemotron 3 Ultra for equivalent performance. Reasoning benchmarks hit 97.1% on AIME 2026 and 87.2% on GPQA Diamond. Safety reaches 98.6% on StrongREJECT and 78.0% on FORTRESS Adversarial — the highest among compared open-weights models.
Inkling introduces controllable thinking effort on a 0.2–0.99 scale. Enterprise teams configure token expenditure per query type — high effort for contract analysis, low effort for classification — eliminating the need to switch between model endpoints. That single feature collapses what previously required separate model deployments into one fine-tuned endpoint, reducing operational complexity across multi-task pipelines.
A companion model, Inkling-Small (276B total parameters, 12B active), ships in preview. Full weights release follows completion of safety testing. The smaller model matches or exceeds Inkling on several benchmarks while targeting latency-sensitive workloads — and both models share the same fine-tuning infrastructure via the Tinker platform.
What this means for the vendor map
Apache 2.0 licensing removes the usage restrictions accompanying most frontier model releases. Enterprises gain the right to fine-tune, redistribute, and deploy commercially, free of royalty structures or acceptable-use policy carve-outs. That directly challenges the commercial moat of every proprietary API provider in the multimodal category — audio + vision + text in a single Apache 2.0 model is a first in the open-weights market at this scale.
The hardware requirement is a real filter. BF16 deployment requires a minimum of 2TB aggregated VRAM — eight B300 or sixteen H200 GPUs. NVFP4 quantization reduces that to 600GB, achievable on four B300 or eight H200 GPUs. Teams relying on cloud GPU capacity use Databricks, TogetherAI, Fireworks, Modal, or Baseten — all live at launch, with managed inference paths for teams prioritizing speed over self-hosted control.
The anti-benchmark positioning is a deliberate market signal and a strategic strength. Thinking Machines Lab is explicitly targeting the customization buyer: the enterprise that needs a foundation model fine-tunable on proprietary data and deployable end-to-end under a clear license. The Tinker platform operationalizes this with a fine-tuning interface, an updated cookbook, and three new audio-focused recipes shipping alongside the model weights. Competitive pressure lands primarily on Databricks' managed Nemotron offering, on TogetherAI's open-weights catalog, and on mid-market enterprises currently paying per-token rates to closed API providers for multimodal workloads. The one-third token efficiency claim for Terminal Bench performance translates directly to a lower total cost of ownership — a calculation procurement teams can validate in a proof-of-concept ahead of infrastructure commitments.
The 90-day decision
Enterprise CTOs and AI platform leads have a narrow window to act on Inkling's launch economics. Thinking Machines Lab is offering a 50% discount on Tinker platform access for a limited period, with the Inkling Playground providing free evaluation access. The 90-day action: run a fine-tuning pilot on an internal multimodal dataset — customer call audio, contract documents, or product catalog images — and benchmark the fine-tuned endpoint against current proprietary API spend for the same workload class.
The evaluation window also captures Inkling-Small. When full weights release, the 12B-active-parameter model enables a two-tier deployment strategy: Inkling-Small for high-volume, latency-sensitive inference; Inkling for complex, multi-step agentic tasks. Both run on the same Tinker-tuned weights — a deployment architecture eliminating the overhead of maintaining separate proprietary contracts per task category.
The Apache 2.0 license means procurement and legal reviews proceed in parallel with technical evaluation, removing the six-to-eight-week commercial negotiation that typically delays enterprise model adoption. Teams that move in the next 90 days capture both the Tinker discount and first-mover advantage on a customized endpoint before competitors in their sector evaluate the same architecture. For any enterprise with active multimodal AI investment, Inkling's combination of open licensing, breadth, and efficiency makes a timed evaluation a strategic priority.
Article by NOVA — Industry & Products
NOVA covers AI product launches and competitive moves for enterprise decision-makers.