Thinking Machines Lab’s Inkling AI model is a 975-billion-parameter, natively multimodal mixture-of-experts system released with downloadable Apache 2.0 weights. It is not the fastest route to the top of every benchmark—and the company explicitly says it is not the strongest model overall. Inkling’s more interesting proposition is breadth: text, image and audio reasoning; controllable test-time effort; a one-million-token ceiling; calibration-focused post-training; and a fine-tuning and deployment stack designed to let developers make the model their own.
That combination makes Inkling a consequential open-weight release, but not an uncomplicated one. The hardware requirements are substantial, most benchmark results remain vendor-reported, and Thinking Machines’ own launch materials disagree on several scores. This analysis uses the dated July 14 comparison table in the Hugging Face model card as its main benchmark reference and documents the discrepancies below.
Inkling AI model: key takeaways
- Large but sparse: Inkling has 975 billion total parameters, with 41 billion active for each token.
- Natively multimodal: It accepts text, images and audio and produces text. Video appears in the training-data description, but it is not listed as a supported launch input.
- Open weights, not a complete open-source training release: The weights carry an Apache 2.0 license, while the full training corpus and an end-to-end reproducible training stack have not been published.
- Competitive, not dominant: Inkling is strong on instruction following, tool use and several safety evaluations, but Kimi K2.6, GLM 5.2, DeepSeek V4 Pro and proprietary frontier systems lead many reasoning and coding tests.
- Accessible in code, expensive in hardware: The BF16 checkpoint needs at least 2 TB of aggregate GPU memory; even the NVFP4 version calls for roughly 600 GB.
What Thinking Machines released
Thinking Machines introduced Inkling on July 15, 2026 as its first open-weight foundation model. It is available as a BF16 checkpoint and an NVFP4 checkpoint, through the company’s Tinker fine-tuning service, and through a group of third-party inference providers.
The model is intended as a broad base for coding agents, tool-using systems, retrieval-augmented applications, chatbots and multimodal workflows. That framing matters. Inkling is not positioned as a narrowly optimized reasoning model. Thinking Machines calls it a generalist and openly acknowledges that stronger models exist.
A second model, Inkling-Small, was previewed with 276 billion total parameters and 12 billion active. Thinking Machines says the smaller model matches or exceeds the large version on several internal evaluations while reducing cost and latency. As of July 15, however, its weights were not present alongside the two public Inkling repositories in the Thinking Machines Hugging Face account. It should therefore be treated as a preview, not a shipped open-weight model.
Inkling specifications at a glance
| Specification | Inkling |
|---|---|
| Provider | Thinking Machines Lab |
| Release date | July 15, 2026 |
| Model type | 66-layer, decoder-only multimodal autoregressive transformer |
| Architecture | Sparse mixture of experts (MoE) |
| Parameters | 975B total; 41B active per token |
| Expert routing | 6 of 256 routed experts, plus 2 shared experts |
| Attention | Sliding-window and global layers at a 5:1 ratio; 8 KV heads |
| Maximum context | Up to 1M tokens; Tinker launch options are 64K and 256K |
| Inputs | UTF-8 text, pixel-based images and 16 kHz WAV audio |
| Output | UTF-8 text |
| Pretraining | Reported 45T tokens across text, images, audio and video |
| Published checkpoints | BF16 and NVFP4; the website model card also lists MXFP8 support |
| License | Apache 2.0, with a separate model acceptable-use policy |
The official Inkling model card recommends image dimensions between 40 and 4,096 pixels and audio under 20 minutes for best performance. Audio must be supplied as 16 kHz WAV. These are operating recommendations rather than absolute statements about every deployment framework.
Architecture and training
Inkling’s MoE design broadly follows the DeepSeek-V3 pattern. A router selects six of 256 specialist feed-forward experts for each token, while two shared experts remain active. The point of this sparse design is to hold much more learned capacity than a 41B dense model while avoiding the compute cost of activating all 975B parameters on every step.
The attention stack departs from today’s most common recipe. Thinking Machines interleaves five sliding-window layers for every global-attention layer, uses relative positional embeddings instead of rotary positional embeddings, and inserts short convolutions after key and value projections and along residual branches. The company says relative position handling extrapolated better to long sequences in its experiments. That is a provider finding, not yet an independently replicated conclusion for Inkling.
Vision and audio are brought into the same decoder space as text. Images are divided into 40-by-40-pixel patches and passed through a lightweight hierarchical patch encoder. Audio is represented with discrete dMel spectrogram tokens. This is significant because multimodality is part of the base model rather than a collection of separate external encoders bolted onto a text-only system.
Thinking Machines reports pretraining on 45 trillion tokens drawn from public sources, third parties and synthetic or augmented data. It does not publish an itemized corpus. For optimization, large matrix weights used Muon while other parameters used Adam. An initial supervised phase used synthetic outputs from open-weight models including Kimi K2.5, followed by the larger investment: more than 30 million asynchronous reinforcement-learning rollouts across math, code, tool use, audio, vision, chat and safety environments.
Controllable reasoning is Inkling’s clearest differentiator
Inkling can be asked to use different levels of reasoning effort. At a low setting, it spends fewer generated tokens; at a high setting, it devotes more computation to the problem. This is more useful than a single “thinking mode” switch because developers can tune the latency, cost and accuracy trade-off for each workload.
Thinking Machines’ effort-sweep charts cover Terminal-Bench 2.1, Humanity’s Last Exam and IFBench. The company says Inkling can match Nemotron 3 Ultra on Terminal-Bench while using roughly one-third as many generated tokens. That is an efficiency claim from the launch evaluation, not an independent cost study, but it points to a practical advantage: a model used millions of times may be valuable because of the shape of its performance curve, not merely its maximum score.
The launch also demonstrates Inkling writing and running its own Tinker fine-tuning job to learn a constrained writing behavior. It is a useful illustration of the intended workflow—model, training API and evaluation loop working together—but it should not be mistaken for evidence that autonomous self-improvement is solved.
Inkling benchmark results
The table below uses the Inkling Hugging Face model card’s July 14 comparison values. Thinking Machines says Inkling ran at effort 0.99 and temperature 1.0, with coding trajectories capped at 256K tokens. Some comparator scores were taken from Artificial Analysis; others came from the company’s own harnesses or the cited benchmark providers. Those differences limit perfect apples-to-apples interpretation.
| Category | Evaluation | Inkling | What it tests |
|---|---|---|---|
| Reasoning | HLE, text only | 30.0% | Broad expert-level questions |
| Reasoning | HLE, with tools | 46.0% | Expert questions with tool access |
| Math | AIME 2026 | 97.1% | Competition mathematics |
| Science | GPQA Diamond | 87.9% | Graduate-level scientific reasoning |
| Agentic coding | SWE-bench Verified | 77.6% | Real repository issue resolution |
| Agentic coding | SWE-bench Pro Public | 54.3% | Harder professional software tasks |
| Agentic coding | Terminal-Bench 2.1 | 63.8 | Terminal-based task completion |
| Tool use | MCP Atlas | 74.1% | Use of MCP-connected tools |
| Tool use | Tau 3 Banking | 22.3% | Multi-step banking-agent workflows |
| Factual research | BrowseComp with context management | 77.1% | Web research and evidence synthesis |
| Factuality | SimpleQA Verified | 43.9% | Short factual answers |
| Instruction following | IFBench | 79.8% | Following precise constraints |
| Vision | MMMU Pro, Standard 10 | 73.3% | Multidisciplinary visual reasoning |
| Vision | CharXiv RQ with Python | 82.0% | Chart understanding with code tools |
| Audio | MMAU | 77.2% | General audio understanding |
| Audio | VoiceBench | 91.4% | Spoken instruction and voice tasks |
| Safety | FORTRESS, adversarial | 78.0% | Refusal of harmful weapon-related requests |
| Safety | FORTRESS, benign | 95.9% | Avoiding over-refusal of safe look-alikes |
| Safety | StrongREJECT | 98.6% | Refusal of unambiguously harmful prompts |
Thinking Machines assigns a score of zero to Terminal-Bench rollouts where web search contaminated the solution. The launch post also warns that one Humanity’s Last Exam chart used an earlier Inkling checkpoint. These are good disclosures, but they reinforce why the numbers should be read as evidence about a specific harness and configuration, not permanent model rankings.
Inkling versus Kimi, GLM, DeepSeek and proprietary models
Inkling’s own comparison table shows a balanced model rather than a category winner. It beats Kimi K2.6, GLM 5.2 and DeepSeek V4 Pro on IFBench in the listed runs, while Nemotron 3 Ultra is slightly higher. Inkling also performs well on MCP Atlas and FORTRESS. On raw frontier reasoning and coding, however, the leaders are usually elsewhere.
| Benchmark | Inkling | Nemotron 3 Ultra | Kimi K2.6 | GLM 5.2 | DeepSeek V4 Pro | Best listed closed model |
|---|---|---|---|---|---|---|
| HLE, text only | 30.0% | 26.6% | 35.9% | 40.1% | 35.9% | Claude Fable 5: 53.3% |
| SWE-bench Verified | 77.6% | 70.7% | 80.2% | — | 80.6% | Claude Fable 5: 95.0% |
| Terminal-Bench 2.1 | 63.8 | 56.4 | 71.3 | 82.7 | 64.0 | GPT 5.6 Sol: 89.5 |
| MCP Atlas | 74.1% | 44.7% | 68.1% | 77.8% | 73.2% | Claude Fable 5: 83.3% |
| IFBench | 79.8% | 81.4% | 76.0% | 73.3% | 76.5% | Gemini 3.1 Pro: 77.1% |
| CharXiv RQ with Python | 82.0% | — | 86.7% | — | — | Gemini 3.1 Pro: 89.9% |
| FORTRESS, adversarial | 78.0% | 77.6% | 65.6% | 71.3% | 36.0% | Claude Fable 5: 96.0% |
The practical reading is straightforward. For maximum general reasoning or terminal coding performance, GLM 5.2, Kimi K2.6 and the strongest closed systems are ahead in this table. For teams prioritizing an Apache-licensed, multimodal base with configurable reasoning and a polished fine-tuning pathway, Inkling occupies a different position.
For a broader view of the field, see Kingy AI’s comparison of the leading open-weight models in 2026, its analysis of Kimi K2.6’s agentic coding focus, and the deep dive into DeepSeek V4 and million-token context economics.
The official benchmark discrepancies
Thinking Machines’ sources do not present one perfectly consistent scorecard. The launch article and website model card list HLE text-only at 29.7%, GPQA Diamond at 87.2%, GDPVal-AA v2 at 1,238, Tau 3 Banking at 23.7%, AA Omniscience at 2.1 and MMMU Pro at 73.5%. The Hugging Face table—labelled as comparison scores generated July 14—lists 30.0%, 87.9%, 1,233, 22.3%, 1.0 and 73.3%, respectively.
The provider does not explain whether these differences reflect checkpoint changes, reruns, rounding or updated external results. They are generally too small to reverse the article’s conclusions, but they matter for reproducibility. This article uses the Hugging Face values for all head-to-head comparisons and preserves the discrepancy rather than blending the tables.
There is a second documentation mismatch: the website model card lists BF16, MXFP8 and NVFP4 numerical support, while the Hugging Face README lists BF16 and NVFP4. Developers considering MXFP8 should confirm support in their selected runtime. Finally, although video is included in the training-data description, the released model’s listed inputs are text, images and audio. “Trained on video” should not be read as a launch claim for direct video input or video generation.
Why the Apache 2.0 weights matter
With a closed API, customers receive outputs under a provider’s pricing, rate limits, moderation and deployment policies. Downloadable weights allow an organization to operate the model inside its own infrastructure, fine-tune it on proprietary tasks, inspect behavior across controlled test suites, choose an inference stack and retain more control over data location and system updates.
Apache 2.0 is a familiar, commercially permissive software license. That lowers legal friction for research and product integration compared with licenses containing usage tiers or field-of-use restrictions. Teams should still review Inkling’s separate acceptable-use policy and obtain their own legal advice rather than assuming the license answers every deployment question.
Inkling is nevertheless better described as open weight than fully open source. Thinking Machines describes the categories and provenance of its data, several architectural choices and portions of the training recipe, but it does not provide the full 45-trillion-token corpus or everything needed to reproduce the released checkpoint from scratch. The weights enable adaptation and study; they do not make the entire development process transparent.
That distinction is central to Inkling’s impact. It expands the number of high-capability multimodal systems that organizations can own and modify, increases pressure on closed providers, and gives researchers another substantial base for post-training work. It does not remove the economic concentration created by frontier-scale training and inference.
Hardware, deployment, fine-tuning and price
The published deployment requirements put “local” in perspective. The BF16 checkpoint needs at least 2 TB of aggregate VRAM, with suggested configurations of eight NVIDIA B300 GPUs or 16 H200s. The NVFP4 checkpoint lowers the requirement to at least 600 GB: four B300s in W4A4 mode, which requires SM100-class hardware, or eight H200s in W4A16 mode.
Supported deployment paths include SGLang, vLLM, TokenSpeed, Unsloth/llama.cpp and Hugging Face Transformers. Hosted API partners named at launch include Together, Fireworks, Modal, Databricks and Baseten. The broad day-one runtime support is one of the release’s strongest practical features. Kingy AI has separately covered how Hugging Face models are moving into managed fine-tuning and deployment workflows.
For customization, Tinker offers Inkling at 64K and 256K context lengths. As of July 15, its models and pricing page lists a limited-time 50% Inkling discount. At 64K, the posted per-million-token rates are $1.87 for prefill, $0.374 for cached prefill, $4.68 for sampling and $5.61 for training. The 256K option lists $3.74, $0.748, $9.36 and $11.23, respectively. Those are Tinker service prices, not estimates for self-hosting, and they are likely to change.
Safety, calibration and known limitations
Thinking Machines groups calibration, instruction following and resistance to censorship under “epistemics.” The model was trained with rubric graders and a factuality grader that uses agentic web search, plus abstention-aware rewards intended to teach it when to answer, hedge or decline to guess. Forecasting evaluations show Inkling in the same broad range as several frontier systems, although the company notes that those results used an earlier checkpoint.
For safety, the provider evaluated harmful compliance, over-refusal, CBRN and cyber uplift, sycophancy, manipulation, vulnerable-user interactions, strategic deception, sabotage and loss-of-control scenarios. It reports Inkling materially below public frontier models on loss-of-control capability and says the release does not add material dangerous-capability uplift beyond what is already available in the open-weight ecosystem.
The model card also documents residual failures. Inkling can comply with some harmful requests framed indirectly or as role-play. Thinking Machines recommends defense in depth, including application-layer filtering, monitoring, rate limits and tools such as Llama Guard. It also warns about hallucinations, imperfect instruction following, degraded performance in long multi-turn conversations and uneven results across languages, cultures and domains.
Those limitations are especially important because weights can be fine-tuned in ways that weaken built-in safeguards. Inkling should not be deployed for medical, legal or safety-critical decisions without domain-specific validation and human oversight.
How Inkling pushes the AI ecosystem forward
Inkling does not push the frontier forward by winning every leaderboard. Its contribution is architectural and economic: it packages multimodal understanding, a large sparse model, controllable inference effort, long context, calibration work and an open-weight customization path into one release.
That matters for three reasons. First, it gives developers an additional non-proprietary foundation for applications that need audio and vision as well as text. Second, it makes reasoning effort an explicit product control rather than a hidden provider decision. Third, it links the downloadable checkpoint to real fine-tuning and serving infrastructure from the first day, reducing the integration gap that often follows model releases.
The counterweight is accessibility. A 600 GB quantized checkpoint remains far beyond ordinary workstation hardware. Inkling broadens freedom for well-resourced research groups and enterprises more than it democratizes frontier AI for individual developers. Inkling-Small could change that equation somewhat, but its 12B active parameters do not reveal the total memory footprint, and the weights had not shipped at publication time.
Final assessment: who should use Inkling?
Inkling is most compelling for organizations that want to customize a multimodal foundation model, control deployment and data handling, and tune reasoning effort to a task. It is also an interesting research base for reinforcement learning, calibration, agentic tools and multimodal post-training.
It is less suitable for teams seeking a model that runs on a single workstation, buyers who only need inexpensive hosted inference, or workloads where the absolute best current reasoning and coding scores matter more than ownership. In those cases, a smaller open model or a frontier proprietary API may be the better engineering choice.
The release deserves attention precisely because the honest verdict is more nuanced than “new number one.” Inkling is a capable, unusually broad and commercially usable open-weight model with serious infrastructure requirements and incomplete reproducibility. Its biggest test will be what developers can create after fine-tuning—not where its launch checkpoint sits on a transient leaderboard.
Frequently asked questions
What is the Inkling AI model?
Inkling is Thinking Machines Lab’s 975B-parameter multimodal mixture-of-experts model. It accepts text, image and audio inputs, produces text, and activates 41B parameters per token.
Is Inkling open source?
Its weights are downloadable under Apache 2.0, so “open weight” is accurate. The complete training corpus and a fully reproducible end-to-end training stack are not published, so describing the entire project as fully open source would overstate the disclosure.
Can Inkling run locally?
It can be self-hosted, but not on ordinary consumer hardware. The published minimums are 2 TB of aggregate VRAM for BF16 or about 600 GB for NVFP4.
Is Inkling better than Kimi K2.6 or GLM 5.2?
Not across the board. Inkling leads some listed instruction-following and safety comparisons, while Kimi K2.6 and GLM 5.2 are ahead on several reasoning and agentic coding tests. Harness and tool differences also limit direct ranking.
Is Inkling-Small available?
Thinking Machines previewed a 276B-total, 12B-active model, but its full weights were not listed in the company’s Hugging Face account as of July 15, 2026.
Editorial note: Kingy AI did not independently run Inkling for this launch analysis. Benchmark values are attributed to the cited provider and evaluation sources, and the discrepancies among Thinking Machines’ official materials are documented above.