Thinking Machines Inkling Multimodal Model Challenges the Closed-Model Default

Thinking Machines released Inkling on July 15, pairing three input modalities with open weights and a context window reaching one million tokens. The Thinking Machines Inkling multimodal model accepts text, images, and audio, then produces text responses through one shared reasoning system.

The release matters because Inkling is not trying to win every benchmark. Thinking Machines explicitly says it is not the strongest open or closed model available. Instead, the company is betting that control, customization, and efficient reasoning can outweigh a narrow performance lead.

That puts Inkling against the closed-model default established by OpenAI, Anthropic, and Google. Those providers offer highly capable systems through managed interfaces, but customers cannot inspect, host, or directly modify their underlying weights.

Inkling offers a different bargain. Developers receive an Apache 2.0 model, downloadable checkpoints, native multimodal processing, and built-in support for post-training through Thinking Machines' Tinker platform.

The model remains demanding to operate, and most performance evidence comes from its creator. Still, this is more than another downloadable language model. Inkling connects open weights with a commercial customization system, deployment partners, and a clear argument about who should control specialized AI.

Thinking Machines Inkling Arrives With Open Weights

Inkling combines a very large sparse architecture with unusually broad developer access.

According to the official Inkling release, the model contains 975 billion total parameters. It activates 41 billion parameters while processing each token.

This design is called a mixture-of-experts transformer. Instead of using every parameter for every request, the model routes each token through a limited group of specialized computational components.

Inkling has 256 routed experts and two shared experts in each expert layer. Six routed experts process each token, while the shared experts remain active across inputs.

That routing makes the difference between the model's total and active parameter counts important. Inkling retains broad capacity without applying all 975 billion parameters during every inference step.

The model was pretrained on 45 trillion tokens spanning text, images, audio, and video. Thinking Machines says it trained Inkling from scratch on NVIDIA GB300 NVL72 systems.

Post-training covered reasoning, mathematics, coding, tool use, images, audio, conversation, and safety. The company also reports using more than 30 million reinforcement-learning rollouts, meaning completed model interactions used to improve behavior through feedback.

Inkling accepts text, pixel-based images, and WAV audio sampled at 16 kHz. The model returns text rather than generated images or speech.

That distinction keeps the product focused on multimodal understanding. Inkling can transcribe speech, interpret recordings, answer questions about images, analyze charts, and combine those inputs with tools.

Thinking Machines lists a maximum context window of one million tokens. Context is the information a model can consider during one interaction, including instructions, documents, tool results, and earlier messages.

Access varies by environment. Tinker currently exposes 64,000-token and 256,000-token options, while the downloadable model supports the larger stated maximum.

Developers can download full weights through the Inkling model repository. The repository identifies the license as Apache 2.0 and provides BF16 and NVFP4 checkpoint options.

The Apache license allows broad use, modification, and distribution, subject to its conditions. That makes Inkling more adaptable than an API-only model whose provider controls the underlying checkpoint.

Thinking Machines also released integrations for common inference systems. Supported options include vLLM, SGLang, TokenSpeed, Unsloth, llama.cpp, and Hugging Face Transformers.

API access is available through infrastructure providers including Together AI, Fireworks, Modal, Databricks, and Baseten. This distribution strategy reduces the need for each developer to assemble a hosting stack independently.

Inkling is therefore open-weight, but not limited to self-hosting. Teams can download it, fine-tune it through Tinker, or use managed inference from an outside provider.

That flexibility creates the article's central tension. Closed providers sell access to finished intelligence, while Thinking Machines wants developers to treat the model itself as editable infrastructure.

Native Multimodal Reasoning Is the Main Technical Bet

Inkling processes text, vision, and audio inside one decoder instead of treating them as loosely connected external services.

Many products described as multimodal rely on separate components. A speech recognizer converts audio into text, an image encoder summarizes a picture, and a language model handles the final reasoning.

Inkling takes a more integrated route. Its modalities are projected into a shared hidden space and processed jointly by the same decoder.

Images are divided into 40-by-40-pixel patches and passed through a hierarchical encoder. Audio enters through dMel representations, which convert speech signals into compact tokens suitable for model processing.

A lightweight embedding layer transforms those inputs before the main transformer handles them alongside text tokens. This architecture lets the model connect evidence across modalities during reasoning.

That capability matters for tasks where transcription alone loses information. Tone, background sounds, diagrams, layouts, and visual relationships can all affect the correct interpretation.

Consider a recorded product interview accompanied by screenshots. A multimodal model can connect a participant's spoken complaint with the interface element visible at that moment.

A developer might also provide an error screenshot, a spoken explanation, and a code repository. Inkling can analyze those inputs in one workflow before proposing a fix or using tools.

The same approach supports document-heavy knowledge work. Teams often handle meeting recordings, charts, technical documents, and written messages as separate information streams.

A searchable knowledge base becomes more useful when AI can reason across those formats. However, dependable retrieval and source verification remain necessary around any generative model.

Thinking Machines reports competitive results on several multimodal evaluations. At its highest tested effort setting, Inkling scored 73.5 percent on MMMU Pro and 78.1 percent on CharXiv RQ.

The company reports 91.4 percent on VoiceBench, 77.2 percent on MMAU, and 56.6 percent on Audio MC. These evaluations test different aspects of speech understanding, audio reasoning, and instruction following.

The numbers should not be treated as a single verdict. Each benchmark uses different prompts, graders, harnesses, and operating conditions.

Thinking Machines acknowledges one concrete limitation in its VoiceBench procedure. The benchmark uses strict string matching, so the company added a system instruction encouraging the required answer format.

Inkling can also use Python for image operations such as cropping and zooming. This gives the model a way to inspect small visual details instead of relying entirely on its first image representation.

The model's multimodal strengths do not mean it produces every media type. Its outputs remain text, so applications requiring generated audio or images need additional models.

Audio also carries practical limits. The official Inkling model card recommends WAV input sampled at 16 kHz and recordings under 20 minutes for optimal performance.

Those constraints place the current release closer to a reasoning engine than a complete real-time voice product. Applications still need interfaces, storage, retrieval, moderation, and output systems around it.

Thinking Machines has connected Inkling to its separate interaction-model research, which explores real-time collaboration through voice and vision. Inkling is intended to become a background reasoning component for that broader system.

The strategic bet is clear. The company expects future AI applications to mix modalities continuously, not send every visual or audio signal through isolated preprocessing steps.

Controllable Thinking Changes the Cost and Performance Tradeoff

Inkling lets developers vary reasoning effort instead of accepting one fixed balance between speed, token use, and answer quality.

Reasoning models often improve when allowed to generate more intermediate computation. That improvement carries costs because longer outputs consume more tokens and increase response time.

Inkling exposes a controllable effort setting from 0.2 to 0.99. Developers can select lower effort for routine requests and reserve higher effort for harder problems.

Thinking Machines trained this behavior by changing system messages and adjusting the per-token cost during reinforcement learning. Different samples therefore rewarded different levels of computation.

The company says Inkling's reasoning traces became more concise as training progressed. Later traces removed grammatical filler while preserving the reasoning needed for the final answer.

That result supports a practical product argument. Efficient reasoning is not only about shrinking the model or quantizing its weights.

A team can also control how much work the same model performs for each task. Classification, summarization, and formatting may need less effort than debugging or scientific reasoning.

Thinking Machines reports that Inkling matched Nemotron 3 Ultra on Terminal Bench 2.1 while using roughly one-third as many generated tokens. Terminal Bench evaluates agents performing tasks inside terminal environments.

This remains a company-reported comparison. The tested models may use different harnesses, prompts, and default operating settings.

Inkling's broader benchmark record is mixed, which aligns with the company's own positioning. The model scored 77.6 percent on SWE-bench Verified and 54.3 percent on the public SWE-bench Pro evaluation.

It reached 46 percent on Humanity's Last Exam with tools and 29.7 percent without them. On GPQA Diamond, it recorded 87.2 percent.

Several competing models beat Inkling on those evaluations. Claude Fable 5, GPT 5.6 Sol, Gemini 3.1 Pro, GLM 5.2, and Kimi K2.6 each lead it in selected categories.

Inkling also trails Gemini on all three audio benchmarks reported in the release. On MMMU Pro, its 73.5 percent result sits below the listed scores for Kimi K2.6, Gemini, Claude, and GPT.

That does not invalidate the model's purpose. It shows why Thinking Machines emphasizes the entire cost curve rather than the maximum score at one operating point.

The strongest model for a single difficult prompt may not be the best choice inside an agent that makes hundreds of calls. Latency and token use compound across planning, tool execution, revision, and evaluation loops.

Developers can fine-tune Inkling for narrow behavior, then run that version at a calibrated effort level. This offers two control surfaces: what the model specializes in and how much computation it spends.

Inkling-Small extends the same argument. The preview model has 276 billion total parameters and 12 billion active parameters.

Thinking Machines reports that the smaller version approaches Inkling on several tasks. It slightly exceeds the larger model on GPQA Diamond, IFBench, MCP Atlas, and CharXiv RQ with Python.

Inkling remains stronger on Terminal Bench, SimpleQA, and several other evaluations. The company has not yet released the full Inkling-Small weights because testing is still underway.

If those results hold across independent evaluations, the smaller model could become more relevant for high-volume workloads. Coding graders, synthetic-data generators, and background agents value predictable cost and response time.

For now, the larger Inkling demonstrates the mechanism. Sparse activation, adjustable effort, and post-training work together to challenge the assumption that one fixed inference profile fits every request.

Open Weights Put Pressure on Closed AI Platforms

The main competition is not Inkling against one model. It is editable model infrastructure against provider-controlled intelligence.

Closed-model services simplify adoption. A developer sends a request to an API, receives an answer, and leaves infrastructure management to the provider.

That arrangement also concentrates control. The provider decides when models change, which behaviors are allowed, how data flows through the service, and which customization methods are available.

Open weights shift several of those decisions downstream. Organizations can inspect deployment behavior, run private infrastructure, preserve a chosen checkpoint, and create specialized versions.

Inkling adds a notable commercial layer to that freedom. Thinking Machines is not simply placing a model online and leaving developers to manage every later step.

Tinker offers supervised fine-tuning and reinforcement-learning workflows through an API. The updated Tinker cookbook includes recipes for adapting Inkling, including examples focused on audio.

The company demonstrated this connection by asking Inkling to modify its own behavior. The model wrote a fine-tuning job, created an evaluation, ran training, and loaded the resulting checkpoint.

The requested behavior was a lipogram, meaning the model had to respond without using the letter "e." Thinking Machines reports that the complete workflow finished in about 27 minutes.

That demo does not establish general self-improvement. It shows that an agent can operate the tooling needed to create and activate a narrow fine-tune.

The distinction matters. A model changing one constrained writing behavior is far less ambitious than independently improving its overall reasoning or correcting unknown weaknesses.

Even so, the workflow captures Thinking Machines' product thesis. The valuable unit is not a static chatbot response but a model that a developer can repeatedly adapt, test, and deploy.

This puts indirect pressure on OpenAI, Anthropic, and Google. Their closed models often lead Inkling on the evaluations listed by Thinking Machines, but customers depend on provider-defined customization boundaries.

Open-weight competitors face more direct pressure. Inkling's architecture follows important elements of DeepSeek-V3, including sparse expert routing and auxiliary-loss-free balancing.

The DeepSeek-V3 paper helped establish a widely studied recipe for efficient mixture-of-experts training. Inkling adopts that foundation while adding its own attention design, multimodal inputs, and controllable effort.

Inkling must also compete with the Qwen, Kimi, GLM, DeepSeek, and Nemotron families. Several already offer strong open-weight reasoning, coding, or multimodal capabilities.

Thinking Machines is differentiating through the combination rather than a single benchmark lead. Inkling joins native audio and vision, long context, open checkpoints, post-training access, and managed deployment partners.

That bundle could appeal to enterprises with specialized data or workflows. A financial firm may need a model adapted to its review criteria, while a software company may optimize for its repositories and tools.

The economics still favor managed APIs for many teams. Hosting a model of Inkling's size requires expensive infrastructure and experienced operators.

The BF16 checkpoint needs at least two terabytes of aggregate GPU memory, according to the model card. Suggested configurations include eight NVIDIA B300 GPUs or 16 H200 GPUs.

The NVFP4 checkpoint lowers the minimum to 600 gigabytes. It can run on four B300 GPUs in one supported mode or eight H200 GPUs in another.

Those requirements make "downloadable" different from "easy to run locally." A laptop cannot host the full model simply because the weights are available.

Managed inference partners close part of that gap. Yet using them reintroduces an outside service, even when the customer retains more choice over the underlying model and fine-tune.

Inkling therefore pressures closed platforms at the level of control, not convenience. Its success depends on whether enough buyers value checkpoint ownership and customization to accept added operational complexity.

The Verification Gap Is Inkling's Biggest Constraint

Inkling's specifications are concrete, but its quality, safety, and efficiency claims still need broader independent testing.

Most benchmark results in the announcement come from Thinking Machines. The company says it uses externally reported scores for competitors when available and internal harnesses for other evaluations.

Cross-model comparisons become difficult when evaluation conditions differ. A stronger agent harness can raise coding scores even when the underlying model remains unchanged.

Thinking Machines says Inkling used a bash-only harness for SWE-bench Verified. Its Terminal Bench result came from an internal coding harness.

The company also identified some Terminal Bench solutions contaminated by web search and assigned those attempts a score of zero. That disclosure is useful, but it illustrates how evaluation details influence final rankings.

Several forecasting results came from a checkpoint different from the released model. Testing occurred between June 30 and July 13, according to the announcement.

The clearest conclusion is therefore narrower than the marketing story. Inkling appears competitive across many tasks, but available evidence does not establish universal superiority.

Thinking Machines avoids that overclaim itself. Its announcement states that Inkling is not the strongest model available, whether open or closed.

Safety presents a similar verification problem. The company reports internal evaluations and external testing covering chemical, biological, radiological, nuclear, cyber, and loss-of-control risks.

On FORTRESS Adversarial, Inkling scored 78 percent. It recorded 95.9 percent on the benign portion and 98.6 percent on StrongREJECT.

Those figures describe refusal behavior under specific tests. They do not guarantee safe conduct after fine-tuning or during novel attacks.

The model card identifies residual susceptibility to role-play and indirectly framed harmful requests. It recommends multiple defenses instead of relying only on the model's refusals.

That advice is especially important for open weights. A downstream developer can weaken, remove, or accidentally damage safeguards during customization.

Thinking Machines says Inkling did not provide material dangerous-capability uplift beyond what is already available in the open-weight ecosystem. Independent researchers still need to test that conclusion.

The training-data description is also broad. Data came from public sources, third parties, and synthetic or augmented material, but the release does not provide a detailed dataset inventory.

This limits outside analysis of copyright exposure, demographic representation, language coverage, and contamination. Training on 45 trillion tokens makes provenance questions more significant, not less.

The model card warns about hallucinations, instruction-following failures, and declining performance during long, multi-turn conversations. It also notes uneven performance across languages and subject areas.

The one-million-token window deserves similar scrutiny. Maximum context capacity does not establish accurate recall or reasoning throughout that entire span.

Long-context models can lose details, overweight recent material, or struggle to combine distant evidence. Independent "needle" tests and realistic document workflows will show how Inkling behaves near its limit.

Native multimodality creates additional failure modes. Speech recognition errors, ambiguous sounds, visual hallucinations, and conflicts between modalities can all distort later reasoning.

Organizations should test the exact inputs they expect to use. A high benchmark score cannot substitute for evaluation on domain-specific accents, diagrams, documents, or safety requirements.

Inkling's Apache 2.0 license broadens experimentation, which should accelerate this verification. Researchers can run controlled tests without waiting for API access or worrying about silent checkpoint changes.

The evidence gap is not a reason to dismiss Inkling. It is the main reason to treat the release as the beginning of an evaluation cycle rather than its conclusion.

Three Signals Will Show Whether Inkling Matters

Inkling's importance will be determined by independent evidence, real customization, and the performance of its smaller sibling.

The first signal is third-party evaluation under consistent conditions. Researchers need to compare Inkling with open and closed alternatives using the same prompts, tools, sampling settings, and compute limits.

Independent multimodal tests matter most. Inkling's distinctive claim involves reasoning across text, images, and audio, not merely matching another model on mathematics.

Strong outside results would support Thinking Machines' broad-model strategy. Large gaps between company and independent scores would weaken the efficiency and quality argument.

The second signal is downstream fine-tuning activity. The model repository already supports adapters, quantizations, and third-party integrations, but quantity alone says little about practical value.

The useful evidence will come from specialized models that outperform their base on real tasks. Examples could include audio analysis, coding agents, document review, forecasting, or domain-specific tool use.

Developers should also watch whether fine-tuning preserves safety and general capabilities. A narrow improvement that damages reliability elsewhere creates a costly deployment tradeoff.

Tinker is central to this test. If teams can repeatedly build, evaluate, and export useful checkpoints, Thinking Machines will have a stronger position than a company offering weights alone.

The third signal is the full release of Inkling-Small. Its 12 billion active parameters make it more plausible for latency-sensitive or high-volume workloads.

Company results suggest that Inkling-Small already approaches the larger model on multiple evaluations. Independent confirmation would strengthen the case for training quality over raw active scale.

A delayed release, weak outside testing, or major capability losses would shift attention back toward the 41-billion-active-parameter model. That would also preserve Inkling's demanding infrastructure profile.

Together, these signals will answer the central question. Can an editable multimodal foundation compete through control and specialization, even when closed models retain higher peak scores?

For developers, the immediate action is straightforward. Test Inkling against a real workflow, use a fixed evaluation set, measure token use and latency, and verify every safety assumption after customization.

For enterprise buyers, the decision is less about one leaderboard. It concerns where model control should live, how much infrastructure the organization can support, and whether specialized performance justifies that work.

The Thinking Machines Inkling multimodal model has made its wager visible: open weights, native multimodality, and adjustable reasoning can form a credible alternative to closed intelligence.

Now the burden moves outside the company. Independent evaluators, fine-tuning teams, and production deployments must show whether that combination delivers dependable advantages beyond the launch benchmarks.