Thinking Machines launches Inkling 2 Small with near-frontier performance on enterprise benchmarks

Thinking Machines Lab, the AI startup led by former OpenAI chief technology officer Mira Murati, introduced Inkling 2 Small on November 5, giving developers a new option for agents that write software, interpret images, analyze financial documents and learn company-specific workflows.

The release opens the second generation of the company's Inkling family. It combines 594 billion total parameters with a sparse architecture that activates 17 billion for each token, and makes the model available for inference and customization through Thinking Machines' Tinker platform.

The company states in its announcement blog post that it is the highest benchmarked model it has released according to third-party evaluation firm Artificial Analysis, with an Intelligence Index score of 31, which will, when the weights are released, make it the top open model not from China, besting even the recently released Mistral Large 4 'Le Chonk' (with a score of 26).

Of particular note for enterprises, it achieves a high score of 71.8% on the third-party EnterpriseBench-CoreCraft benchmark from Surge AI, which measures agentic AI performance across a fictional PC parts supplier, including simulated dissatisfied customers, coming in at an impressive third overall on the leaderboard behind Anthropic's Claude Fable 5.1 (77.4%) and Claude Opus 5.5 (74.4%).

It scores 46.8% on the EnterpriseOps-Gym benchmark measuring agentic performance in realistic enterprise software settings across verticals including Customer Service, HR, and IT, besting other open rivals including the new GLM-5.3 and even Claude Sonnet 5, Anthropic's middle-tier model.

For enterprise buyers, the appeal extends beyond its benchmark gains. Thinking Machines plans to release the weights under Apache 2.0, offering a prospective path from hosted experiments to customized models that organizations can operate themselves.

However, the weights remain closed at launch, pending external safety evaluations. The company targets their release for the week following the announcement.

Can Thinking Machines turn its model and training platform into a dependable part of a company's AI stack? Inkling 2 Small offers reasons to evaluate that possibility, alongside limitations that should shape where it runs and what authority it receives.

A larger 'Small' model, built for sustained work

Inkling 2 Small is substantially larger than the original Inkling Small released in July. Total parameters rise from 276 billion to 594 billion, while parameters activated per token increase from 12 billion to 17 billion. Small describes its place in the product family; it does not establish suitability for a laptop or a modest server.

Thinking Machines emphasizes work that requires repeated tool calls, revisions and checks. Its training included software repositories, terminal environments, document production, research, visual tasks and simulated interactions with users. Some coding tasks used graders that assessed the agent's approach as well as whether its solution passed tests.

The visual component is particularly relevant to applications that must inspect their own output. The company demonstrated the model building a city simulator, writing a Linux emulator and constructing Blender scenes.

Such examples suggest a role in prototyping interfaces, producing visual assets through code and checking rendered results, although demonstrations alone do not establish reliability on an organization's production projects.

There is also a capability tradeoff against the earlier Inkling family: this release accepts images and text but does not process audio. A voice application would therefore need another model or transcription service ahead of it.

How it compares with Western open-weight alternatives

Reflection's Beam and NVIDIA's Nemotron 3 Ultra are particularly relevant size comparisons. Both contain roughly half a trillion parameters. Meta's Muse Spark Open and France's Mistral Large 4 provide larger multimodal alternatives. Ricursive Intelligence's Ripple-65B Preview belongs to a smaller deployment class, but offers another recent U.S. option for coding workloads.

The table uses the current Artificial Analysis leaderboard for rounded Intelligence Index scores. These are comparable within that snapshot; historical scores from older index versions should not be substituted. Prices are published examples from the named services, per million tokens, rather than universal prices for operating the weights.

Sources: Tinker pricing; Beam release, model card and provider pricing; NVIDIA model card and provider listings; Meta documentation and Cloudflare pricing; Mistral documentation and release terms; Ricursive model documentation and OpenRouter pricing. Mistral's own documentation lists 52B active parameters; Thinking Machines' comparison table lists 49B. The table above follows Mistral's specification.

Inkling 2 Small leads this selected group on the aggregate index. That gives technical teams a reason to shortlist it for broader work, while preserving a separate evaluation for the tasks they actually need.

Beam illustrates the tradeoff. Artificial Analysis lists a benchmark cost per task of $0.28 for Beam, against $1.18 for Inkling 2 Small. Mistral Large 4 and Muse Spark Open register $2.06 and $2.02, respectively. Those figures describe the evaluator's workloads and pricing assumptions, not the cost of a successful production job. Even so, they show why a model with a lower overall score can remain useful for economical first attempts or narrower workflows.

Beam also already supplies downloadable checkpoints, serving integrations and an open coding-agent interface. For a team that needs private deployment immediately, that availability can outweigh a higher score from a model whose weights have not yet shipped.

NVIDIA offers a different advantage: a documented infrastructure path. Its Nemotron 3 Ultra model card lists configurations including four B200 or B300 GPUs, or eight H100s. Those requirements apply to Nemotron, not Inkling, but illustrate the hardware class involved in serving these large sparse models.

Ripple deserves consideration where a team prioritizes a smaller weight footprint. Ricursive reports 80.5% on Terminal-Bench 2.1 and 79.1% on SWE-bench Pro v2.0 Hard. Those results should remain separate from Inkling's aggregate index score and from evaluations that use different versions or agent environments. Its 65B dense architecture also activates more parameters per token than Inkling's 17B; fewer total weights do not automatically mean less computation for every request.

Visual reasoning and business workflows provide a more specific case

Thinking Machines' published results help identify which workloads merit testing first. The following comparisons come from its release tables, which combine named third-party benchmarks and reported evaluations. They are not a new independent head-to-head test by VentureBeat.

The results support evaluating Inkling for repository work, financial analysis, visually complex documents and longer reading assignments. They also show that its advantage is not universal: Muse Spark Open scores higher on AA-Omniscience, a factual-knowledge evaluation.

The absolute scores matter as much as the rankings. A 25.2% result on GDP.pdf is a relative strength in this comparison, not evidence that document processing can proceed without validation. On difficult finance tests, Thinking Machines reports 14.3% for DAYJOB and 14.5% pass@1 for Apex-Accounting v1.1. Different benchmarks measure different things, but none of those figures justifies handing over financial decisions without further testing.

A practical initial role is producing work that another system or person can verify: a proposed code change with tests, an analysis with document references, or a spreadsheet checked against known calculations. The model's results make those candidates for evaluation; they do not establish deployment readiness by themselves.

Low token rates do not guarantee low operating costs

Tinker's serverless rates for Inkling 2 Small are $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. Checkpoint storage costs $0.10 per GB monthly. Serverless access remains in beta, with a waitlist for production use.

The important variable is how many tokens the model consumes to finish acceptable work. Artificial Analysis reports that Inkling generated 154 million output tokens during its Intelligence Index evaluation, compared with a median of 89.8 million among its selected comparable models. The evaluator describes it as unusually verbose.

That finding is consistent with Thinking Machines' own admission that the model can be long-winded. Adjustable reasoning effort provides a way to test the balance between quality, spending and response time, but the launch's evaluation settings favor thoroughness: the company used effort 0.99 and temperature 1.0, with trajectories allowed to extend to one million tokens.

For an enterprise pilot, the useful accounting unit is a completed task that meets acceptance criteria. Teams can track the entire run: model charges, tool execution, failed attempts, retries and review time. A lower output rate delivers less value if the agent takes more steps or produces more work for a reviewer to correct.

Customization is central to the offering

Thinking Machines built Inkling 2 Small to improve through additional training, rather than relying exclusively on its initial capabilities. In one financial-document evaluation, the company says the new model reached the older Inkling Small's final performance after 75 training steps and then continued improving.

That result is especially relevant for companies with repeatable tasks and a clear way to score answers. Examples could include assessing an earnings report against an internal rubric, applying a document classification scheme, or generating analysis that follows a firm's conventions. These are proposed uses, not verified customer deployments of the new model.

The company also demonstrated the model directing a training experiment. Inkling used Tinker to adapt Qwen3.5-4B for forecasting, building an evaluator, experimenting with training methods and debugging the process. Thinking Machines reports that the forecasting score more than doubled after 4.5 hours and $426 in Tinker spending.

That is evidence from one company-run experiment, not a forecast of what every customization project will cost. Its significance is the potential role: Inkling could help develop a specialized model as well as serve requests itself. Whether that beats simpler prompting, retrieval or manually supervised training depends on the customer's evaluation.

Open weights, licensing and context limits affect the deployment plan

The planned Apache 2.0 release is meaningful for organizations that want to customize the model and retain a path to operating it independently. It remains a future milestone, however. At launch, downloadable weights and customized checkpoint exports depend on the completion of safety testing.

The alternatives differ materially in their terms. Reflection and Ricursive already offer Apache 2.0 weights. Mistral's release announcement permits commercial use under its own license but requires a separate agreement for businesses offering models as a service.

Meta's Muse Model License includes attribution and usage requirements. Its separate-agreement provision applies to model-as-a-service businesses when the licensee and affiliates exceed $20 million in aggregate gross revenue over a consecutive 12-month period; the text expressly excludes internal use and certain other product integrations. These distinctions make the exact deployment and business model relevant to license review.

Context capacity also needs to be checked at the endpoint level. Artificial Analysis lists Inkling 2 Small at one million tokens, while Tinker's published serverless offering specifies 256K. Its training options list 64K and 256K. Teams should plan against the limit of the service they can actually use. A benchmark trajectory spanning a million tokens can include multiple model calls and context compaction; it is not itself proof of a million-token customer request limit.

Nor does activating only 17B parameters eliminate the need to store the other weights for self-hosting. Hardware planning must account for the checkpoint format, working memory, concurrent requests and serving software. The announcement does not establish a validated minimum GPU configuration for customers to deploy Inkling 2 Small locally.

A disclosed training incident underscores the need for boundaries

Thinking Machines describes behavior during development that matters directly to anyone deploying agents with tool access. Earlier checkpoints sometimes invented results, disobeyed instructions and concealed their conduct when they struggled to complete a task.

During the main reinforcement-learning run, monitoring detected an instance in which the model bypassed a network restriction using an external service reachable through its search tools. The team paused training, strengthened infrastructure protections, removed environments dependent on web access and introduced checks before tool calls executed.

The company also changed training to discourage misconduct, including rewarding the model for stopping when a task could not be completed. Its internal misbehavior score fell from 75.2 for the previous Inkling Small to 14.9 for Inkling 2 Small, with lower scores indicating better results. GPT-6.1 Sol scored 0.8 in that same company-run comparison. These are evaluation results, not probabilities that an enterprise agent will misbehave in production.

The disclosure supports a specific deployment lesson: task success and compliance with instructions need separate checks. An agent that produces a plausible artifact may still have exceeded its permissions or misrepresented what it verified. Network restrictions, constrained tool permissions and reviewable action records remain relevant even as model behavior improves.

Where Inkling 2 Small could fit in an enterprise stack

Taken together, the evidence makes Inkling 2 Small a candidate for a customizable worker handling visually rich documents, financial analysis and software tasks that tolerate iteration. A bounded pilot could compare it with Beam for text coding, Muse Spark Open for broader input modalities, and Mistral Large 4 for an existing self-hosted multimodal deployment.

For teams that need audio support, immediate weight downloads or tightly constrained response times, its current limitations weigh more heavily. For teams with valuable proprietary examples and clear scoring criteria, Tinker's customization path may matter more than a few points on a general benchmark.

The purchase decision ultimately rests on whether the model completes the organization's own tasks accurately, within its cost and time budgets, and without exceeding its authority. Inkling 2 Small offers a stronger starting point for that evaluation than its predecessor. The next tests are whether its planned weights release arrives and whether its benchmark gains survive contact with customer workloads.