Thinking Machines Lab has released an open-weight model that matches or beats its much larger sibling on several coding and reasoning tests. The trade is real: Inkling-Small costs less to run, but gives up a troubling amount of factual recall.
The answer in one sentence: Inkling-Small is the more practical Inkling for coding agents, multimodal experiments and custom fine-tuning, provided teams can tolerate weaker factual recall and either rent the model or supply at least 180GB of aggregate VRAM.
An AI model with 276 billion parameters has no business sounding small. Thinking Machines Lab calls it Inkling-Small anyway, and within this particular family the name is defensible: its first Inkling model has 975 billion parameters.
The useful number is smaller still. Inkling-Small activates 12 billion parameters while processing each token, compared with 41 billion for Inkling. Thinking Machines says that reduction produces comparable performance at roughly one-quarter the size. The arithmetic is closer to 28 or 29 percent, depending on whether total or active parameters are compared. A later, smaller model has recovered most of the larger model’s capability and surpassed it on several coding, reasoning and instruction-following evaluations.
That result does not make Inkling-Small an across-the-board upgrade. Its SimpleQA Verified score is 20.6 percent, less than half Inkling’s 43.9 percent. Its result on the τ³-Banking agent benchmark also falls well behind. “Comparable” describes a new balance of strengths, weaknesses and cost, not interchangeable behaviour.
For developers choosing a base model to adapt, that balance may matter more than a clean leaderboard victory. Inkling-Small is Apache 2.0 licensed, accepts text, images and audio, supports adjustable reasoning effort and can be fine-tuned through Thinking Machines’ Tinker platform. The full weights are downloadable. Running them is another matter.
Quick verdict
| Best for | Skip if |
|---|---|
| Coding agents and tool-use systems where cost per completed task matters | You need strong unaided factual recall |
| Teams planning domain-specific fine-tuning through Tinker | You expect a “small” model to run on an ordinary laptop or gaming GPU |
| Multimodal document, chart, speech and audio experiments | Your deployment cannot absorb beta-service or self-hosting complexity |
| Researchers and companies that need commercially usable open weights | You need production guarantees without doing your own evaluation and safeguards |
Price checked: July 30, 2026. Tinker’s serverless beta lists Inkling-Small at $0.30 per million input tokens, $0.06 for cached input and $1.20 per million output tokens. Promotional training prices and third-party hosting prices differ.
Testing scope: Kingy reviewed the announcement, model card, released repository, current pricing documentation and independent Artificial Analysis results. Kingy did not run the model, reproduce the benchmarks or measure latency, throughput, factuality or hardware use.
What Thinking Machines released
Inkling-Small is a 42-layer, decoder-only Mixture-of-Experts transformer. It contains 256 routed experts; each token is sent to six of them, while two shared experts remain active. That sparse routing lets the model hold 276 billion parameters without executing all of them for every token.
According to the official model card, the released checkpoint supports a context window of up to one million tokens. It accepts UTF-8 text, pixel-based images and 16kHz WAV audio, then generates text. Images are divided into 40-by-40-pixel patches and audio is represented as dMel spectrograms before both are processed with text in the model’s shared hidden space.
The weights are available in the public, ungated Hugging Face repository under Apache 2.0. Thinking Machines lists BF16, MXFP8 and NVFP4 numerical support. At launch, the main BF16 repository contained 32 model-weight shards, while a separate quantized checkpoint offered a lower-memory deployment path.
Developers can approach the model three ways:
- Download and operate the weights on suitable infrastructure.
- Fine-tune it through Tinker, with 64K and 256K training contexts priced separately.
- Use Tinker’s 256K serverless inference beta or its text, image and audio playground.
The downloadable model is documented for up to one million tokens, while Tinker serverless and the configuration evaluated by Artificial Analysis expose 256K. The weights and hosted routes do not share one context limit.
Why the smaller model caught the larger one
Inkling-Small began training after Inkling, which gave the lab time to change its pre-training data mix and machine-learning recipe. Thinking Machines also says an earlier preview checkpoint was post-trained in part through on-policy distillation, using Inkling as the teacher. The lab then ran another two weeks of agentic coding reinforcement learning.
That sequence matters more than the family naming. The larger model helped produce training signals for the smaller one, while the later model benefited from a revised recipe and a more mature post-training stack. Inkling-Small is therefore a cleaner test of how much performance a lab can recover through better training rather than additional parameters.
It recovered a lot.
The benchmark result is strong, but not clean
The following figures come from Thinking Machines’ published evaluation table. The company says some results were obtained through its own harnesses and others were sourced from named external evaluators. They are evidence of reported performance, not a Kingy reproduction.
| Evaluation | Inkling-Small | Inkling | What it suggests |
|---|---|---|---|
| SWE-Bench Verified | 80.2% | 77.6% | Stronger reported coding-agent performance |
| Terminal-Bench 2.1, best harness | 64.7% | 63.8% | Near parity in the company’s harness |
| Humanity’s Last Exam, text only | 31.6% | 29.7% | Better difficult reasoning and knowledge aggregate |
| GPQA Diamond | 89.5% | 87.2% | Small advantage on graduate-level science questions |
| IFBench | 82.2% | 79.8% | Better instruction following |
| Artificial Analysis Index v4.1 | 40 | 41 | Similar composite capability independently |
| τ³-Banking | 15.5% | 23.7% | Clear loss on a general agentic workflow |
| SimpleQA Verified | 20.6% | 43.9% | Major factual-recall weakness |
The coding story survives outside the company’s table, but the exact scores move with the harness. Artificial Analysis reported a score of 40 for Inkling-Small and 41 for Inkling on its Intelligence Index. It also found the models tied at 55 percent on its Terminal-Bench 2.1 setup, rather than reproducing the 64.7 and 63.8 percent figures from Thinking Machines’ best-harness comparison.
Agent benchmarks depend heavily on prompts, tool access, context management and the surrounding software harness. Thinking Machines also says it assigned zero to a small number of Terminal-Bench solutions contaminated by web search. The defensible conclusion is narrow: Inkling-Small operates in roughly the same coding tier as Inkling and can beat it under some reported setups.
Artificial Analysis adds a useful qualification to the knowledge story. Inkling-Small scored minus nine on AA-Omniscience, compared with positive two for Inkling. Incorrect answers therefore outweighed correct answers under that index. Its measured hallucination rate was slightly lower than Inkling’s, 57 percent versus 63 percent, but its accuracy fell from 40 to 31 percent. The smaller model was not merely more reckless. It knew less.
That distinction changes how the model should be deployed. Coding agents can compile, test and inspect their work. Retrieval systems can ground answers in supplied documents. A general-purpose assistant answering from memory has fewer external checks. Inkling-Small looks much better in workflows that can verify the model than in workflows that reward broad, unaided recall.
Efficiency is the product
Inkling-Small’s practical advantage is not a lower sticker price in isolation. It combines fewer active parameters, adjustable reasoning effort and competitive output-token use.
Tinker’s serverless inference pricing lists Inkling-Small at $0.30 per million input tokens and $1.20 per million output tokens. Inkling is listed at $1.00 and $4.05 respectively. Cached input falls to $0.06 for Inkling-Small and $0.17 for Inkling.
At those rates, Inkling-Small’s output is about 70 percent cheaper per token. The active-parameter count is about 71 percent lower. Those ratios will not translate directly into a 70 percent lower application bill because real workloads vary in prompt length, reasoning tokens, retries, tool calls, cache hits and success rates.
Artificial Analysis measured roughly 24,000 output tokens per Intelligence Index task for Inkling-Small and 25,000 for Inkling. That is a modest family-level reduction, but far below the roughly 45,000 it reported for DeepSeek V4 Flash at the same composite score. Inkling-Small finished the AA-Briefcase agent benchmark in an average of 34 turns, compared with 81 for Inkling, although the two models had nearly identical rubric pass rates. Better presentation and shorter trajectories helped the Elo result more than a large increase in correctness.
This is why cost per successful task matters more than token price. A cheap model that reasons in circles can lose its price advantage. Early evidence suggests Inkling-Small avoids some of that excess work.
“Small” does not mean local
The open weights give developers control, but they do not make deployment casual.
Thinking Machines says the BF16 checkpoint needs at least 600GB of aggregate VRAM. Its documented examples are four NVIDIA B300 GPUs or eight H200s. The NVFP4 checkpoint reduces the requirement to at least 180GB. The listed configurations are one B300 in W4A4 mode, which also requires SM100-or-newer architecture, or two H200s in W4A16 mode.
That is smaller than the 975-billion-parameter Inkling, not small by consumer hardware standards. A team can experiment through Tinker or a third-party host without owning the cluster, but then open weights do not remove dependence on an operator. They create the option to move, inspect, customize or self-host when the economics justify it.
Tinker’s serverless inference is also labelled beta. Thinking Machines explicitly says it does not recommend the service for intensive production use yet and asks production customers to join a waitlist. Teams should not mistake public playground access for a mature service-level commitment.
This same tension runs through the broader open-weight model market. Downloadable weights can weaken model lock-in while increasing demand for memory, serving software, security controls and specialist infrastructure. Freedom at the model layer transfers more operational responsibility to the deployer.
Who should use Inkling-Small
Coding and tool-use teams
Inkling-Small has the clearest case here. Its reported SWE-Bench, Terminal-Bench, MCP Atlas and instruction-following results place it close to or above Inkling while its hosted output price is much lower. Teams still need repository-specific evaluations, tool-call reliability tests and cost measurements from complete trajectories.
Organizations planning a custom model
The release is tightly coupled to Tinker. A company can start from commercially usable weights, fine-tune against its own examples and export checkpoints. That path is more consequential than a generic chat comparison, especially for narrow workflows where a customized model can overcome gaps in the base model.
Multimodal builders
Native text, image and audio inputs make Inkling-Small unusual among open-weight models in its active-parameter class. The company reports strong chart, diagram, visual-mathematics, speech and longer-audio performance. Kingy has not independently tested those modalities, and the model card recommends audio clips under two minutes for optimal performance.
Teams with retrieval and verification
The factuality results make grounding essential. Inkling-Small is a more credible fit when it can retrieve source material, call tools, execute code or pass outputs through deterministic checks. High-stakes medical, legal and safety-critical uses require domain validation, application-level safeguards and human oversight, as the model card itself recommends.
What the release says about Thinking Machines
Inkling was the company’s first large training effort. Two weeks later, the smaller sibling arrived with a revised data mix, a revised recipe, teacher distillation and additional coding RL. That is a fast demonstration of learning at the model-family level.
The strategy is coherent. Thinking Machines is releasing weights, offering a customization platform and using its larger model to improve a cheaper base. The value proposition moves from renting a general chatbot toward shaping a model around a particular job.
Open weights also make that claim inspectable. Developers can download the checkpoint, run their own evaluations and move a derivative outside Tinker. Apache 2.0 permits commercial use. Training data and the full training pipeline are not published, so “open weight” remains more accurate than “fully open source.”
The policy case for open-weight AI often centres on access and national competitiveness. Inkling-Small adds a commercial version of the argument: a model does not need to win every benchmark if it is efficient, adaptable and portable enough to become somebody else’s specialized system.
Final verdict
Inkling-Small is a large, infrastructure-heavy open-weight model that extracts an unusual amount of coding, reasoning and multimodal capability from 12 billion active parameters. It is cheaper to sample, easier to adapt and at least as capable as Inkling on several workflows where outputs can be tested.
Its knowledge weakness prevents a blanket recommendation. A 20.6 percent SimpleQA score beside Inkling’s 43.9 percent is not leaderboard noise. Teams using the model as an ungrounded source of facts are choosing the wrong member of the family.
Use Inkling-Small for coding agents, tool-driven workflows, retrieval-grounded applications and domain fine-tuning. Use the full Inkling, a stronger factual model or a routed model stack when broad recall matters more than inference efficiency. With that division understood, Inkling-Small is the more practical release.
Inkling-Small FAQ
What is Inkling-Small?
Inkling-Small is a general-purpose open-weight multimodal model from Thinking Machines Lab. It is a sparse Mixture-of-Experts transformer with 276 billion total parameters, 12 billion active parameters and native text, image and audio input.
Is Inkling-Small open source or open weight?
Open weight is the more precise term. The model weights are available under the commercially permissive Apache 2.0 licence, but Thinking Machines has not published the complete training dataset and reproducible training pipeline normally associated with fully open-source development.
Can Inkling-Small run locally?
It can be self-hosted, but not on ordinary consumer hardware under the configurations documented by Thinking Machines. The BF16 checkpoint requires at least 600GB of aggregate VRAM, while the NVFP4 checkpoint requires at least 180GB. Hosted access is available through Tinker and third-party providers.
How does Inkling-Small compare with Inkling?
Inkling-Small uses less than one-third of Inkling’s total and active parameters and matches or beats it on several coding, reasoning and instruction-following benchmarks. Inkling remains much stronger on SimpleQA and other knowledge-focused evaluations. The best choice depends on whether efficiency and customization or broad factual recall matters more.
Kingy Launch Brief
Put the week’s verified AI launches in your inbox.
One source-checked edition every Friday, with a clear try, watch or skip verdict. After subscribing, check your inbox and confirm your address.
Free · Fridays · Double opt-in · Unsubscribe anytime