Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026, describing a multimodal sparse mixture-of-experts model with 594 billion total parameters and 17 billion active per token. The launch is real; an open-weights release is not. At announcement, access was limited to Tinker, while full weights were described as a future plan conditional on external safety testing and red-teaming.
What launched—and when
Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, in a company post titled “Introducing Inkling 2 Small.” Its model card gives the same release date. Although this story is being considered in a November 6 news cycle, the company’s own dated materials say November 5. That precision matters because the release’s initial availability and promised next steps should not be mistaken for completed events.
The company describes Inkling 2 Small as its second model generation, aimed at agentic coding and research, visual interpretation and creation, knowledge work, reasoning and customization. It accepts text and image input and generates text output. Thinking Machines says it is intended for general-purpose and developer applications, including tool-using systems, coding assistants and retrieval-augmented applications. Those are intended uses, not independent guarantees of reliability across every workflow.
There is a concrete change in distribution status behind the headline. The launch post says users can try the model through Tinker Playground and use Tinker for inference and fine-tuning. But it explicitly characterizes launch access as “closed weights.” The announcement’s promise to publish the full weights under Apache 2.0 “next week” was conditional on external safety testing and red-teaming. As of November 9, that language remains a forward-looking commitment in the cited announcement, not proof that weights have already been released.
A large model with sparse activation
The headline parameter count needs two numbers. Thinking Machines reports 594 billion total parameters and 17 billion active parameters. In a mixture-of-experts (MoE) design, a routing mechanism selects some expert networks for a given token rather than activating every parameter in the model on every step. The active count therefore describes the portion used for an individual token under the stated architecture; it does not mean the system is a 17-billion-parameter model in the ordinary dense-model sense.
The model card supplies additional architectural detail: a 48-layer, decoder-only transformer with a sparse MoE feed-forward backbone, routing each token to eight of 384 experts, and a hybrid of local and global attention layers. The card describes images as encoded through a vision transformer and projected into a shared hidden space with text for joint processing. It also lists MXFP8 and NVFP4 numerics support and a context window of up to one million tokens. These are provider-supplied specifications; the public independent article confirms the headline size and image/text input but does not separately audit the full implementation.
“Small” in the model’s name should consequently not be read as a description of total parameter count, hardware needs, price or latency. Sparse activation can change compute characteristics, but total deployed cost depends on the serving system, memory, routing, batching, precision and the number of generated tokens. Neither the name nor the 17B active figure alone establishes that the model is inexpensive to run.
What independent testing says
Artificial Analysis, an independent model evaluator, published a release analysis dated November 5. It reports an Artificial Analysis Intelligence Index score of 31 for Inkling 2 Small and says this result matches Google DeepMind’s Gemini 3.8 Flash (high) and exceeds Xiaomi’s MiMo-V2.6-Pro on the evaluator’s cited index. Artificial Analysis characterizes the model as proprietary at launch, with weights planned for a later date. Its analysis is useful independent evidence about benchmark performance and launch status, but it is not an audit of the training data, safety process or every product claim made by the developer.
Artificial Analysis reports a number of task-specific results. It gives Inkling 2 Small 25% on FrontierCode 2.0, 13% on Terminal-Bench 4.0, 56% on CritPt v2, and a 46% result on AutomationBench-AA. It says the model’s score on the Intelligence Index is up 18 points from the previous Inkling Small, but also notes substantially higher output-token use: approximately 150,000 output tokens per index task versus about 37,000 for Inkling Small. These figures come from the evaluator’s own testing and reporting, and should not be treated as universal results for all prompts, settings or user applications.
The independent report also shows why a single headline score can conceal trade-offs. In its comparisons with Claude Sonnet 5.5 (medium), Inkling 2 Small does better on FrontierCode 2.0 in the report’s figures—25% versus 11%—but scores 13% on Terminal-Bench 4.0 compared with 30% for Claude Sonnet 5.5 (medium). The report says Inkling 2 Small’s AA-Omniscience hallucination rate is 27%, down from 61% for Inkling Small, while its attempt rate falls from 77% to 53% and its accuracy rises only from 32% to 34%. Artificial Analysis interprets the lower hallucination rate largely in the context of more selective answering: fewer attempted answers are part of the result, not a footnote to omit.
Scores need context
There is a small but material difference between the score displayed in the release materials. Thinking Machines’ announcement says the model achieved 31 on the Artificial Analysis Intelligence Index. The model card’s evaluation table lists 30.6 in a row labeled “AA Index (v5.0).” Artificial Analysis’s November 5 article rounds its reported Intelligence Index result to 31 and identifies its index as v5.0.1. The most responsible reading is that the headline 31 is a rounded evaluator result and the card shows a more precise figure under a version label; the public materials do not fully explain the 0.1-version labeling difference, so these should be reported with their source/version rather than falsely presented as identical records.
Index rankings are comparative summaries of a defined evaluation suite, not a direct measure of general intelligence or a prediction that a model will outperform another system in every task. Benchmarks can differ in task mix, scoring rules, prompting, tool access, reasoning settings and sampling. For example, Thinking Machines’ release post reports its own evaluation tables across coding, vision and knowledge work; those company-run claims are informative but not interchangeable with Artificial Analysis’ independently run index and task scores.
In its company post, Thinking Machines says Inkling 2 Small scores 81.6% on MMMU Pro (Standard 10), 25.4% on FrontierCode v2.0, and 86.3% on SWEBench Pro v2 (Hard), alongside several other results. The official model card’s larger comparison table includes results for numerous model variants, with a note identifying “Inkling 2 Small: Pi harness” for one Terminal Bench entry. Such labels matter: harness, model setting and task version can change scores. A number without the benchmark version and setup is not enough for a fair comparison.
Access, weights and safety
At announcement, Tinker was the available route. The official post says users could try the model in Tinker Playground and use the platform for fine-tuning and sampling. The model card likewise describes distribution through Tinker for inference and fine-tuning, and gives Tinker Terms of Use as the listed license at that stage. This is a hosted service arrangement, not a public checkpoint that users can download, inspect or independently deploy.
The company says it intends to make full weights available under Apache 2.0 after external safety testing and red-teaming, targeting “next week” relative to its November 5 post. That is meaningful as a stated plan, but the condition and future tense must remain attached to the claim. A planned permissive license is not yet an operative license for weight files, and a target timetable is not confirmation of a release. The source materials cited here do not establish that the promised weights had appeared by November 9.
The announcement frames this sequence as a safety process and links to the company’s explanation of a safe path to open weights. Readers should distinguish a described process from independently verified outcomes. The release page does not publish a detailed external red-team report, identify its complete scope, or supply a date-stamped sign-off in the announcement itself. Until those artifacts and the actual files are available, it is not possible to assess whether the process is complete or reproduce the model’s behavior outside Tinker.
Why the distinction matters
For developers, a hosted launch can enable early experimentation without providing the control associated with open weights. Tinker users may evaluate the model on tasks and integrate it through the platform, but the platform’s terms, availability, pricing and service limits shape that use. The independent analysis lists token prices of $0.30 per million input tokens, $1.00 per million output tokens and $0.03 per million cached input tokens, and estimates $1.18 per Artificial Analysis Intelligence Index task. Those are evaluator-reported prices and task-cost estimates at publication, not a promise of cost for every workload.
High output-token consumption deserves attention in agentic settings, where a model may take many steps, inspect files or produce extended reasoning and tool interactions. Artificial Analysis estimates roughly four times as many output tokens per index task as the predecessor. That does not prove every real-world request will use four times as many tokens, but it warns buyers that benchmark quality and efficiency are separate dimensions. Cost, speed and reliability should be tested on representative workloads before teams commit.
For the open-model ecosystem, the immediate news is therefore a capable model available through one provider’s hosted system, accompanied by a conditional commitment to release weights. If fulfilled, an Apache 2.0 weight release could enable broader deployment and third-party examination. Until then, claims about local hosting, modification, independent safety review and replicability remain prospective. The gap between “announced” and “available to download” is not semantics: it defines what users can do today.
What remains unknown
The official launch and model card establish provider-reported architecture, capabilities, date and distribution; Artificial Analysis adds a credible independent performance perspective. The available evidence still leaves open questions. The sources do not prove that an open checkpoint has been published by November 9, detail all conditions of future access, or allow outsiders to reproduce the proprietary Tinker evaluation environment. They also do not settle how the model will compare on a buyer’s own datasets, languages, safety policies or latency requirements.
Accordingly, the defensible headline is not that Thinking Machines has released an open model. It has launched Inkling 2 Small for hosted use and announced a conditional plan to release full weights later. The model’s scale is 594B total and 17B active parameters; the independent evaluator reports a rounded score of 31, with specific strengths and weaknesses across benchmark families. Both the technical result and its distribution status deserve that qualification.
FAQ
When did Thinking Machines announce Inkling 2 Small?
The official release and model card are dated November 5, 2026. The date should not be shifted to November 6 simply because it is grouped in that day’s news cycle.
Is Inkling 2 Small open weight now?
The launch materials say it was initially available with closed weights through Tinker. The company planned a full-weight release under Apache 2.0 after external safety testing and red-teaming; that future commitment is not evidence the weights are already public.
What do 594B and 17B mean?
594 billion is the reported total parameter count, while 17 billion is the reported active parameter count. The model card describes a sparse MoE system that routes each token to eight of 384 experts.
What independent score did it receive?
Artificial Analysis reports an Intelligence Index score of 31 in its November 5 article, identified with v5.0.1. Thinking Machines’ model card lists 30.6 in a v5.0 row. The versions and displayed precision should be attributed rather than treated as a contradiction-free single number.
Does a 31-point index score mean it is best at every task?
No. The index aggregates performance across a defined suite. Artificial Analysis reports both comparative strengths and weaker results, including different outcomes on FrontierCode and Terminal-Bench. Task-specific validation remains necessary.