On November 5, 2026, Thinking Machines Lab, the startup led by former OpenAI CTO Mira Murati, introduced Inkling 2 Small, the first model of the second Inkling generation. It is aimed at agents that write software, read images, analyze financial documents and learn company-specific workflows. For teams weighing open models against closed APIs, it is a name worth adding to the shortlist.
This article draws on VentureBeat's report published on November 6, 2026 (Pacific time), which relays figures from Thinking Machines and from independent evaluators.
A "Small" model with 594 billion parameters
The name is a bit misleading. Inkling 2 Small has 594 billion parameters in total, up from 276 billion in the original Inkling Small released in July. Thanks to a sparse architecture, only 17 billion parameters are active per token, compared with 12 billion before. "Small" describes its place in the product family, not a size that fits on a laptop or a modest server.
The model is available for inference and customization through Thinking Machines' own Tinker platform. It accepts text and images, but unlike the earlier Inkling family it does not process audio. If you build a voice application, you will still need a separate transcription service or model in front of it.
Strong scores, and limits worth stating plainly
According to the company, citing Artificial Analysis, Inkling 2 Small reaches an Intelligence Index of 31. Once the weights are released, that would make it the top open model from outside China, ahead of Mistral Large 4 at 26.
Two enterprise-oriented results stand out:
- EnterpriseBench-CoreCraft (Surge AI): 71.8%, third on the leaderboard behind Claude Fable 5.1 (77.4%) and Claude Opus 5.5 (74.4%). The test simulates a PC parts supplier and its unhappy customers.
- EnterpriseOps-Gym: 46.8%, ahead of open rivals such as GLM-5.3 and even Anthropic's Claude Sonnet 5. It measures agent performance in realistic enterprise software settings like customer service, HR and IT.
In the company's own comparison tables, the model beats Meta's Muse Spark Open and Mistral Large 4 on several tests, including DeepSWE v1.1 (72.1%), JobBench (70.5%) and Finance Agent v2.1 (67.5%). Muse Spark Open, however, scores higher on AA-Omniscience, a factual-knowledge benchmark, so the lead is not universal.
Absolute numbers deserve a sober read too. Inkling scores 25.2% on GDP.pdf, better than its rivals but far from a level that lets you skip verification. On hard finance tests the company reports 14.3% on DAYJOB and 14.5% on Apex-Accounting v1.1. In short, financial decisions should not be handed to the model without further testing.
Apache 2.0 weights are promised, but not shipped yet
What will interest enterprises most is the plan to release the weights under Apache 2.0, which would allow self-hosting and customization on your own infrastructure. At launch, though, the weights remained closed pending external safety evaluations, with the company aiming to release them the week after the announcement.
If you need private deployment today, options with weights already available, such as Reflection's Beam (Apache 2.0), may be more practical. Beam scores lower (18 on the Artificial Analysis index), but its benchmark cost per task is about $0.28 versus $1.18 for Inkling 2 Small. Those are the evaluator's workloads and pricing assumptions, not the cost of a successful production job, yet they show why a lower-scoring model can still serve cheap first attempts.
Low token prices do not guarantee low running costs
On Tinker, serverless pricing is $0.30 per million input tokens, $0.03 for cached input and $1 for output. Training costs $1.50 per million tokens at 64K context or $3 at 256K, and checkpoint storage is $0.10 per GB per month. Serverless access is still in beta, with a waitlist for production use.
When you budget, count every tool call, revision and re-check an agent makes, not just the price list per token.
Where to start testing Inkling 2 Small
Prioritize work whose output a person or another system can verify:
- A proposed code change accompanied by tests.
- A document analysis that cites locations in the source.
- A spreadsheet checked against known calculations.
- A UI or graphic asset built in code, where the model inspects its own rendered result.
The company's demos, such as a city simulator, a Linux emulator and Blender scenes, show potential, but they do not prove reliability on your real projects. The safer route is a small test suite built from your own data and workflows, then a side-by-side comparison with the models you already use.
The wider market is moving fast. Closed-model vendors keep pushing agent features, as in Claude Personal Finance analyzing spending through Plaid, while also warning about risk, as in Anthropic's warning about nearing self-improving AI. Delaying Inkling's weights for safety review shows that safety now shapes open models too, much like the access limits OpenAI chose for GPT-6 Cyber.
Bottom line: shortlist it, don't swap everything yet
Inkling 2 Small suggests non-Chinese open models are closing the gap with closed systems on enterprise agent tasks. But the weights are not out, the service is in beta and several hard tests still show low scores, so treat it as a candidate for evaluation rather than a default choice.
If your business is considering AI agents, model fine-tuning or an evaluation pipeline before rollout, contact the Besoft team for advice that fits your data and infrastructure.
Sources
- VentureBeat, "Thinking Machines launches Inkling 2 Small with near-frontier performance on enterprise benchmarks", Carl Franzen, Nov 6, 2026: https://venturebeat.com/technology/thinking-machines-launches-inkling-2-small-with-near-frontier-performance-on-enterprise-benchmarks
- Thinking Machines, Inkling 2 Small announcement, Nov 5, 2026: https://thinkingmachines.ai/news/inkling-2-small/
- Artificial Analysis, Inkling 2 Small evaluation: https://artificialanalysis.ai/articles/thinking-machines-releases-inkling-2-small-highest-scoring-open-weights-model-outside-of-chinia-upon-weights-release-it
