Thinking Machines Launches Inkling 2 Small, Weights Next Week. What About Arabic?

Thinking Machines Lab, the startup founded by former OpenAI CTO Mira Murati per The Decoder, announced its second-generation model, Inkling 2 Small, on November 5, 2026. Headline pricing looks aggressive: serverless inference is listed at $0.30 per million input tokens, $0.03 for cached input, and $1.00 for output tokens. However, the catch is immediate: the model launched with closed weights. Despite claiming the “top spot among western open-weight models,” Thinking Machines is restricting access exclusively to its proprietary Tinker platform, holding the weights closed until next week “pending further safety evaluations.”

What You Get Today and What Waits for Next Week

Inkling 2 Small is a Mixture-of-Experts (MoE) model built with 594 billion total parameters and 17 billion active parameters. It accepts text and image inputs, supports a 256K-token context window with compaction for longer trajectories, and features an adjustable “variable effort” setting . The vendor developed the architecture using reinforcement learning on hundreds of thousands of coding tasks, positioning it primarily as a base model for fine-tuning.

Yet the rollout follows a frustrating trend: declaring an open-weight milestone while locking the actual files away. For now, users can only test the system via Tinker Playground or run training and sampling on Tinker’s cloud infrastructure. Thinking Machines stated: “We intend to make the full weights available under an Apache 2.0 license next week” after external safety testing and red-teaming conclude, with availability on other platforms planned only “in the coming weeks.” Until those weights become downloadable on hubs such as Hugging Face, engineering teams cannot host the system on their own clusters or inspect the model themselves.

The Price: Cheap per Token, but Factoring Higher Token Usage

On Tinker’s pricing page, dated October 19, 2026, serverless inference rates look competitive against earlier generation models. However, serverless inference is “currently in beta,” requiring enterprise teams seeking production scale to submit their use cases to an email waitlist.

For fine-tuning and sampling on Tinker, the vendor charges $0.30 per million prefill tokens, $0.03 for cached prefill, $1.00 for sampling, and training rates of $1.50 at 64K context or $3.00 at 256K context. Fine-tuned checkpoints become downloadable only when the weights are released.

However, developers must look past raw per-token numbers. As reported by The Decoder, Inkling 2 Small consumes roughly four times as many output tokens per task as its predecessor, driving up total per-task compute costs. While evaluation by Artificial Analysis notes that overall per-task cost remains well below proprietary giants like Gemini 3.8 Flash and Claude Sonnet 5.5, the higher token volume means your monthly spend will not drop as sharply as headline rates suggest.

Who Should Care—and Who Should Avoid It

The model’s strengths and limitations clearly segment who should pay attention today:

• Agent builders: Developers creating autonomous coding workflows benefit from training focused on long-horizon agent loops, with context compaction across its 256K-token window.
• Document and finance analysts: The vendor claims leading capability in quantitative reasoning across financial filings, spreadsheets, and complex documents, and the model can read images.
• Teams investing in proprietary fine-tuning: The Decoder describes Inkling models as base models meant for fine-tuning. Once released, the promised Apache 2.0 license will provide far greater commercial freedom than the more restrictive open-weight licences from Meta and Mistral, according to The Decoder (see our Mistral Large 4 story).

Conversely, teams looking for a standard conversational assistant should steer clear. Thinking Machines openly acknowledges that Inkling 2 Small is “less optimized for the chat experience.” The vendor notes issues including poor content organization, an annoying overuse of slashes, and an overreaction to mental-health and safety queries. The model has no audio support, and The Decoder notes that it does not always choose optimal tools when working inside large codebases.

The Arabic Question: A 15-Minute Playground Test

For technology teams across the Arab world, the official release materials contain a glaring gap: Arabic is not mentioned anywhere. The announcement states only that pretraining incorporated “a broad mix of natural language, code, and image data.” Thinking Machines provides no list of supported languages, and it could not be confirmed whether Tinker sign-up is instant, whether any free credits are offered, or if regional access restrictions exist.

Rather than relying on unverified claims, developers should spend 15 minutes running three concrete tests inside Tinker Playground:

• Dialect comprehension and MSA synthesis: Prompt the model with an informal customer complaint written in Egyptian or Gulf dialect, asking it to summarize the issue and draft a formal response in Modern Standard Arabic. Check whether it captures regional colloquialisms or defaults to English.
• Arabic visual document parsing: Given the vendor’s claimed strengths in document analysis and visual reasoning, upload a scanned Arabic invoice or a screenshot of an Arabic balance sheet. Instruct the model to compute totals and parse table values to test both its OCR accuracy and handling of Arabic text.
• Right-to-left web layout creation: Ask the model to generate a small web interface in Arabic. Verify whether its generated code properly sets dir=”rtl”, arranges navigation elements from right to left, and maintains coherent Arabic text alignment.

What to Do Now

For organizations in the Arab world, Inkling 2 Small is not ready to replace production conversational interfaces, nor should anyone build mission-critical services around a serverless beta with an email waitlist. Its immediate value is experimental.

The true strategic opportunity depends on next week’s promised Apache 2.0 release. If Thinking Machines delivers on its commitment, regional institutions and enterprises will gain access to an open-weight foundation model with 17 billion active parameters. Unlike closed cloud APIs, an Apache 2.0 release permits local hosting within regional data centers and enables companies to fine-tune the base weights on their own Arabic datasets and internal records. Use Tinker Playground today to test its reasoning and OCR capabilities, but wait for the weights to drop before committing serious engineering resources.