Thinking Machines has released Inkling 2 Small, giving organizations a mixture-of-experts model with 594 billion total parameters and 17 billion active parameters for stronger agentic, coding and visual work. Its Artificial Analysis Intelligence Index score reached 31, an 18-point improvement over Inkling Small, although higher token use raises the reported cost per task to $1.18.
The company designed the model as a base for fine-tuning in specialized applications. It accepts text and image inputs, supports adjustable thinking depth, and is initially available through the Tinker platform, with weights planned for release under an Apache 2.0 license after external safety testing.
What Does Inkling 2 Small Bring to Model Selection?
Inkling 2 Small is the second model generation from Thinking Machines, a startup founded by former OpenAI CTO Mira Murati. Its mixture-of-experts design keeps 17 billion parameters active at a time within a 594-billion-parameter model, allowing it to perform more demanding work without activating the entire model for each task.
According to Thinking Machines, the model scores 31 on the Artificial Analysis Intelligence Index. That is 18 points above Inkling Small, which scored 13, and roughly equal to Google’s Gemini 3.8 Flash on the index.
AutomationBench-AA rose from 5% for Inkling Small to 46%.
- AutomationBench-AA rose from 5% for Inkling Small to 46%.
- FrontierCode 2.0 rose from 5% to 25%.
- AA-Briefcase reached 1,527 Elo, 615 points above its predecessor.
- The reported hallucination rate fell from 61% to 27%, although the model also answered fewer questions.
- The response rate declined from 77% to 53%.
The lower hallucination rate came with fewer answers. The response rate fell from 77% to 53%, so agencies can compare the 27% hallucination rate with the lower response rate when evaluating models for client workflows.
Muse Spark Open and Beam Models in 2026
Inkling 2 Small joins two other Western open-weight releases from 2026 that emphasize coding and agentic workloads. Meta’s Muse Spark Open is a multimodal mixture-of-experts model with approximately 816 billion total parameters and around 42 billion active parameters per token. Meta trained it for agentic coding tasks, including reading codebases, writing and debugging code, and calling tools. It accepts text, images, audio, and video as inputs, produces text output, and supports a context window of up to one million tokens. Its release followed the smaller Muse Glimmer model from the same series.
Reflection’s Beam was also introduced in 2026 as the company’s first open-weight model. It uses a sparse mixture-of-experts architecture with 501 billion total parameters and 23 billion active parameters for coding, reasoning, and agentic workloads. Beam was pretrained on 23.8 trillion curated tokens from the web and proprietary licensed datasets. When announced, it was undergoing final red-teaming and evaluation, with its weights, technical report, model card, and developer artifacts planned for release later that month. Together, these models provide additional open-weight options to consider alongside Inkling 2 Small.
What Does Inkling 2 Small Cost to Run?
Inkling 2 Small generates about 150,000 output tokens per task on average, approximately four times the number generated by its predecessor. Artificial Analysis puts the cost at $1.18 per task, nearly six times the predecessor’s cost.
That increase remains below the reported costs of Gemini 3.8 Flash at $7.77 and Claude Sonnet 5.5 at $10.02 per task. It also places Inkling 2 Small closer to the price range of low-cost Chinese models.
For agencies managing client websites, assess the model’s reported average cost of $1.18 against the value and length of each task. It generates about 150,000 output tokens per task on average, so include output length in the comparison. A model that uses more output tokens can still produce a lower cost per completed task than some alternatives when task completion and fewer retries matter.
How Strong Is the Model at Coding?
Thinking Machines says it trained Inkling 2 Small with reinforcement learning across hundreds of thousands of long-horizon coding tasks. An automated evaluation system checked whether generated code worked and how the model approached each task.
The reported coding results include:
- 25.4% on FrontierCode v2.0
- 72.1% on DeepSWE v1.1
- 86.3% on SWEBench Pro v2 (Hard)
DeepSeek V4.1 Flash still leads on FrontierCode, with 28.5%, and on DeepSWE, with 77%. Kimi K3 narrowly leads the compared field on SWEBench Pro v2 (Hard). The newer Kimi K3.1 Pro was absent from the comparison.
Thinking Machines also acknowledges that the model sometimes communicates inefficiently and does not always choose the best tools for large codebases. That makes it unsuitable as an automatic choice for every coding task without a controlled evaluation.
What Can the Model Do With Images and Documents?
Thinking Machines reports that Inkling 2 Small leads the compared open-weight models in text recognition, infographic interpretation and visual reasoning. On the document benchmark GDP.pdf, which asks questions based on complex PDF files, the model scored 25.2%, the highest result among the compared models.
Additional multimodal results include 81.6% on MMMU Pro and 64.7% on PerceptionBench. It earned a 1,340 Elo rating on DesignArena Website, where people compare AI-generated websites in head-to-head evaluations, also the highest score among open-weight models in the comparison. The company says the model has also made progress in 3D modeling.
In a demonstration, the model worked from photographs in a blind wine tasting and correctly identified three wines in 19 of 20 attempts. The test was a company demonstration rather than a general measure of visual performance.
How Was Inkling 2 Small Trained?
Thinking Machines says the full training run processed about 12 trillion tokens, fewer than comparable rivals. The reinforcement-learning phase after base training used 3 trillion tokens on 2,048 GPUs over nine days.
The model is part of a planned family of different sizes. Thinking Machines also says more variants will follow.
What Limitations Should Agencies Check?
Agencies evaluating Inkling 2 Small for client websites should test the specific workflow that matters, including output length, cost, tool selection and answer completion. The model does not support audio, is less optimized for chat, can overreact to health-related topics and may produce cluttered output.
Those limitations are relevant when a client expects polished conversational responses, audio processing or tightly controlled output. For a specialist coding or document workflow, they may matter less than the higher benchmark results and fine-tuning potential.
What to Do Now
- Review the reported cost per task against the tasks your agency actually runs. The model’s $1.18 average comes with about 150,000 output tokens per task, so output length needs to be included in the evaluation.
- Run a controlled comparison on one representative client workflow. Measure completed tasks, retries, response rate and tool use instead of relying only on a general intelligence score.
- Check the model’s behavior on the content your clients handle. Thinking Machines notes problems with inefficient communication, poor tool selection for some large codebases, health-topic overreaction, cluttered output, chat optimization and audio support.
- Plan for two access stages. The initial offering is through Tinker, while open weights are planned under Apache 2.0 after external safety testing.
- Evaluate specialized fine-tuning as well as the base model. Thinking Machines designed Inkling models as base models for fine-tuning in specialized applications.
FAQ
Who Developed Inkling 2 Small?
Inkling 2 Small was developed by Thinking Machines, the startup founded by former OpenAI CTO Mira Murati.
How Large Is Inkling 2 Small?
It is a mixture-of-experts model with 594 billion total parameters and 17 billion active parameters at any given time.
How Does Inkling 2 Small Compare With Inkling Small?
Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, compared with 13 for Inkling Small. Its reported cost is $1.18 per task, nearly six times the predecessor’s cost, and it generates about 150,000 output tokens per task on average.
Sources
Want help with this in Myrtle Beach or Long Island? Talk to us.
NewSunSEO
Grow in Google and AI Search
AI-SEO, local search, paid ads and websites from one team, with offices on Long Island and in Myrtle Beach.
- AI-SEO
- Local AI-SEO
- AI Consulting
- PPC
- Web Design
- Web Development
- Long Island Office(631) 320-1700
- Myrtle Beach Office(843) 492-7131
Request a Free Proposal
Tell us about your business and where you want to grow.
We will reply by email to talk through your goals and what a proposal would include.
We use your details only to reply to you and send what you asked for. See our Privacy Policy.
