Inkling-Small hits near-flagship scores at quarter size under Apache 2.0
Thinking Machines followed its flagship Inkling release with Inkling-Small, a 276-billion-parameter open-weight model that lands within a single point of the flagship on the Artificial Analysis Intelligence Index while using roughly a quarter of the active parameters. The new model posts 80.2% on SWE-bench Verified against Inkling's 77.6%, and 64.7% on Terminal Bench 2.1 against 63.8%. The 'Small' label is relative, not consumer-scale: the BF16 checkpoint still demands roughly 600 GB of aggregate GPU memory. The pipeline angle and the Apache 2.0 license may be the parts of this release that outlast the benchmark headline.
The benchmark profile is a clear tradeoff rather than a uniform win. Thinking Machines reports Inkling-Small ahead on SWE-bench Verified, Terminal Bench 2.1, SciCode, Humanity's Last Exam, GPQA Diamond and CritPt. The flagship retains a meaningful advantage on factual recall and specific agentic tasks. Inkling-Small scores 15.5% on τ³-Banking against Inkling's 23.7%, and its AA Omniscience score is negative, reflecting weaker factual coverage even with a slightly lower reported hallucination rate. The workload split reads cleanly from the numbers: Inkling-Small matches coding assistants, tool-use systems, retrieval-augmented generation, and multimodal pipelines; high-stakes factual work still demands retrieval, verification, and human review. The source does not specify whether those gaps narrow under the model's variable-reasoning mode or under heavier test-time compute.
The architecture explains how a 276-billion-parameter model can use only 12 billion parameters per inference step. According to the model card, Inkling-Small is a sparse Mixture-of-Experts design: a 42-layer decoder routes each token to six of 256 specialized experts, with two additional shared experts active for every token. The system retains a large pool of learned capacity while activating a fraction of it per step. The model is also natively multimodal, with images, audio, and text projected into a shared representation and processed jointly by the decoder rather than handled through separate external systems. A one-million-token context window is supported, alongside a variable-reasoning-effort control that lets developers trade quality, latency, and cost at the call level. The source does not detail which layer carries the integration cost when ingestion formats or schemas change in production.
The deployment floor is the practical constraint. Thinking Machines lists two supported configurations for the BF16 checkpoint: 4x NVIDIA B300 GPUs or 8x NVIDIA H200 GPUs. A quantized NVFP4 checkpoint lowers the requirement to roughly 180 GB of aggregate VRAM, fitting a single B300 in W4A4 mode or two H200s in W4A16 mode. Ordinary laptops, MacBooks, gaming desktops, and most developer workstations fall outside the supported envelope. The realistic targets are enterprise GPU servers, cloud clusters, and specialized inference providers. The reduction from the flagship is real and meaningful for capacity planning, hosting cost, and self-hosting feasibility, but 'small' is being measured against a 975-billion-parameter sibling, not against the universe of runnable local models.
The Apache 2.0 license may matter as much as the size. Apache 2.0 is one of the most familiar permissive terms in the software industry and generally permits use, modification, fine-tuning, redistribution, and commercialization inside proprietary products, subject to standard notice and attribution requirements. That removes several friction points that custom 'open' licenses often introduce, including revenue thresholds, branding obligations, and additional commercial conditions for large-scale deployment. The source points to Moonshot's recent Kimi K3 weight release under a custom license with separate commercial terms as a contrast case. Apache 2.0 lowers the legal review burden even though it does not by itself resolve acceptable-use policy, data provenance, or downstream safety obligations.
API economics reinforce the deployment argument. The company is advertising a limited-time 50% discount, bringing the standard 64K-context variant to $0.58 per million prefill tokens, $1.44 per million output tokens, $1.73 per million training tokens, and $0.116 per million cached prefill tokens. A 256K-context variant is available at higher rates. Full BF16 and NVFP4 checkpoints are also distributed, and deployment paths run through SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face tooling, in addition to the Tinker API for fine-tuning. The combination gives enterprises a choice between hosted inference, third-party providers, fine-tuning through the company's own service, and private infrastructure operation.
The release pattern carries a separate signal that the benchmarks do not capture. Thinking Machines researcher Horace He contrasted the two launches in a post on X, writing that Inkling-Small 'felt much more routine' and that the team essentially reused the Inkling pipeline with a smaller configuration. The company says Inkling-Small benefited from an improved pre-training data mix, recipe changes, and on-policy distillation using Inkling as a teacher, followed by two weeks of continued agentic coding reinforcement learning. Mira Murati emphasized the same point on X, framing the new release as a quarter-size open-weight sibling with fine-tuning access through Tinker from day one. Read together, the two launches suggest the company is moving toward a cadence in which open-weight multimodal systems can be produced, improved, and released more routinely, rather than treated as isolated research milestones.
Whether teams can verify throughput, consistency, and cost against the hardware configurations they already operate is the deployment question the source does not test. The official benchmarks, license, pricing, and deployment paths are all documented; behavior under concurrent load, irregular delegation graphs, and production-grade schema changes is not. The gap between what these benchmarks measure and what enterprise deployment demands will determine whether the smaller footprint becomes a real operating advantage or just a cheaper line item on a GPU invoice.