Models & Open Source
Inkling-Small Makes Control a 23% Premium
Inkling-Small nearly matches its 975B sibling, but beta hosted inference costs 23% more than GPT-5.6 Luna. Open weights are the product.
Thinking Machines released Inkling-Small with 12 billion active parameters and nearly the benchmark profile of its 975-billion-parameter sibling. For a representative 100-million-input, 20-million-output-token workload, Tinker’s beta serverless rates produce a $54 bill versus $44 on GPT-5.6 Luna—a 23% premium that matters only if open weights, private deployment, or fine-tuning create real value.
Small is a topology, not a laptop promise
The name is relative. Thinking Machines describes Inkling-Small as a 276-billion-parameter mixture-of-experts model that activates 12 billion parameters per token, compared with 975 billion total and 41 billion active for Inkling. That is 28.3% of the total parameter pool and 29.3% of the active footprint. The model accepts text, images, and audio, supports up to a one-million-token context window, and ships with full weights under Apache 2.0.
The benchmark trade is unusually attractive for coding. The company’s model card reports 80.2% on SWE-bench Verified versus 77.6% for Inkling, and 64.7% on Terminal-Bench 2.1 versus 63.8%. Humanity’s Last Exam also favors the smaller model, 31.6% to 29.7%. Thinking Machines continued agentic-coding reinforcement learning for two weeks after improving the training recipe, so “smaller” here also means newer.
But compression did not preserve every capability. Inkling-Small scores 15.5% on the τ³-Banking agent benchmark, well below Inkling’s 23.7%, while SimpleQA Verified falls to 20.6% from 43.9%. The company acknowledges that Inkling retains an advantage in knowledge coverage and factuality. That split suggests a clean routing policy: code and tool use deserve a pilot; high-stakes factual workflows need retrieval, verification, and human review before they earn production traffic.
Harness caveats matter. Thinking Machines says its Terminal-Bench result uses an internal coding harness and that contaminated web-search solutions were zeroed. The model card compares against vendor-reported scores where available, which does not create a perfectly level test. Still, VentureBeat’s launch analysis notes that Artificial Analysis gave Inkling-Small an index score of 40 versus 41 for Inkling. The compression story survives independent aggregate measurement even if no single coding score should settle procurement.
The hardware floor eliminates casual self-hosting. The model card requires at least 600GB of aggregate VRAM for BF16—four B300s or eight H200s. NVFP4 lowers that to 180GB, still one B300 or two H200s. The weights are downloadable from Hugging Face, but downloadability is not affordability. Teams should model server utilization, operations labor, failover capacity, and moderation rather than comparing an owned cluster’s token cost with a fully managed API’s sticker price.
That distinction extends the argument in Kimi K3’s open-weight economics: control has value, but permissive weights do not repeal electricity, scarce accelerators, or on-call duty. Inkling-Small’s Apache 2.0 license removes many commercial-use ambiguities. It does not remove the platform team.
Pay the premium only when it buys an option
The hosted arithmetic clarifies what the product is. Tinker’s serverless beta prices Inkling-Small at $0.30 per million input tokens and $1.20 per million output tokens. Apply those rates to 100 million input and 20 million output tokens: $30 plus $24 equals $54. OpenAI’s GPT-5.6 Luna model page puts GPT-5.6 Luna at $0.20 input and $1.20 output, making the same mix $20 plus $24, or $44. Inkling-Small’s beta bill is 22.7% higher.
That is not a universal total-cost comparison. Tinker explicitly discourages intensive production use of the beta, while Luna is a production API; reasoning output, caching, providers, and task success can alter the denominator. Tinker’s discounted training/sampling tier costs $0.58 per million prefill tokens, $1.44 per million sampled tokens, and $1.73 per million training tokens, but those are not like-for-like hosted inference rates. The defensible conclusion is narrower: Inkling-Small is not the obvious hosted bargain simply because it has “Small” in the name.
The open-weight option is separate from the 23% API premium. Full weights let a regulated company keep inference inside a controlled environment. Apache 2.0 allows modification and commercial redistribution with familiar obligations. Tinker exposes training at a discounted $1.73 per million tokens, creating a path to specialize the model around proprietary tasks rather than perpetually prompt a closed generalist. Those options can outweigh $10 per illustrative beta workload when data residency, model behavior, or vendor exit carries a larger expected cost.
The counterpoint is that most teams will never exercise them. A startup that downloads weights but serves through a third party has gained licensing flexibility, not infrastructure independence. An enterprise that fine-tunes without a durable eval set may pay to make a model confidently idiosyncratic. And a team that self-hosts at low utilization can spend far more on idle H200s than it saves on tokens.
This is why Amazon’s $220 billion capacity build matters to the model decision. Hyperscalers are spending aggressively to turn scarce accelerators into discounted commitments. Open weights improve bargaining power only when the operator can credibly move inference among owned infrastructure, specialist hosts, and clouds.
The practical verdict has three parts. Teams needing private multimodal inference or model modification should test Inkling-Small now, using coding and tool-use tasks where its profile is strongest. Hosted general-purpose workloads should benchmark it against Luna and other cheap models on cost per successful task, not parameter count. High-stakes banking, legal, or factual systems should wait for stronger domain evidence or wrap the model in retrieval and approval controls.
What would change the verdict? Stable, non-promotional Tinker pricing; independent same-harness evaluations; production inference partners with service levels; and utilization data showing that a 180GB deployment can beat managed APIs after labor and redundancy. Until then, Inkling-Small is a compelling control plane disguised as a cheap model. Buy it for the control—not the adjective.