# Inkling 2

> Source: https://aiwiki.ai/wiki/inkling_2
> Updated: 2026-11-17
> Fact-checked: 2026-11-17
> Categories: AI Models, Large Language Models, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Inkling 2." aiwiki.ai, 17 Nov 2026. https://aiwiki.ai/wiki/inkling_2
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Inkling 2** is the second generation of [large language models](https://aiwiki.ai/wiki/large_language_model) from [Thinking Machines Lab](https://aiwiki.ai/wiki/thinking_machines_lab), the company led by former OpenAI chief technology officer [Mira Murati](https://aiwiki.ai/wiki/mira_murati). As of November 2026 the generation has one released member, **Inkling 2 Small**, a sparse [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) model with 594 billion total parameters and 17 billion active per token that accepts text and images and produces text [1][2]. Thinking Machines announced it on November 5, 2026, initially offering it only through its [Tinker](https://aiwiki.ai/wiki/tinker) platform with the weights withheld, and published the weights on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) under the [Apache License 2.0](https://aiwiki.ai/wiki/apache_license) on November 12, 2026, after external safety testing [1][5]. The company describes Inkling 2 Small as "part of a ladder of models" and says more sizes will follow, but it has not named or dated any of them [1]. On the Artificial Analysis Intelligence Index v5.0.1, Inkling 2 Small scored 31, against 13 for Inkling Small and 12 for the original [Inkling](https://aiwiki.ai/wiki/inkling) on the same index version [6][8].

## Background

Thinking Machines released its first model family in July 2026: Inkling, a 975-billion-parameter mixture-of-experts model with 41 billion active parameters, on July 15, and Inkling-Small, with 276 billion total and 12 billion active parameters, on July 30 [16][17][10]. Both took text, image and audio input and were published with Apache 2.0 weights [16][17][19]. At the end of July the company published "A Safe Path to Open Weights," a framework under which access to a model widens in stages, from hosted access for selected defenders through monitored public access to full weights, "only when the evidence supports it" [11]. Inkling 2 Small was released on a staged schedule of that kind, with a week of hosted-only access on Tinker before the weights went out [1].

## Release and licensing

The announcement on November 5, 2026, said the model was available for fine-tuning and sampling on Tinker "with closed weights, following the approach we outlined in a safe path to open weights," and that the company intended to release the full weights under Apache 2.0 the following week, once external safety testing and red-teaming were complete [1]. [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis), writing the same day, described it as "launching as a proprietary model" with weights to follow, and made its ranking of the model among open-weight releases conditional on that release [6]. VentureBeat likewise reported that the weights remained closed at launch [12].

The weights were uploaded to the `thinkingmachines` organization on Hugging Face on November 12, 2026, and an update dated that day was added to the announcement [1][5]. Two checkpoints were published:

| Repository | Format | Files on Hugging Face | Minimum aggregate VRAM (model card) | Weights license |
| --- | --- | --- | --- | --- |
| `thinkingmachines/Inkling-2-Small` | [NVFP4](https://aiwiki.ai/wiki/nvfp4) (native precision) | about 394 GB | 500 GB, e.g. 2x NVIDIA B300 or 4x H200 [2] | Apache 2.0 [2] |
| `thinkingmachines/Inkling-2-Small-MXFP8` | MXFP8 | about 623 GB | 800 GB, e.g. 4x NVIDIA B300 or 8x H200 [2] | Apache 2.0 [4] |

The model was pretrained, midtrained and post-trained in NVFP4, and the model card recommends the NVFP4 checkpoint for all uses; it says the MXFP8 version scores 1 to 2 percentage points higher on evaluations, a gap the company does not consider meaningful in practice [2]. Unlike the first Inkling, whose original checkpoint was in bfloat16, Inkling 2 Small has no bfloat16 checkpoint; the 8-bit MXFP8 version is the highest-precision release [2][18]. Both repositories declare the Apache 2.0 license in their metadata, with a link to the license text at apache.org rather than a separate LICENSE file, and the model card links to the company's Model Acceptable Use Policy [2][4]. As of November 17, 2026, the Hugging Face organization lists no other Inkling 2 checkpoint [19].

## Architecture

The model card describes Inkling 2 Small as a 48-layer decoder-only transformer whose feed-forward layers are a sparse mixture of experts, with each token routed to 8 of 384 experts, and whose attention mixes local and global layers [2]. The configuration file published with the weights gives the details:

| Property | Inkling 2 Small | Inkling (first generation) |
| --- | --- | --- |
| Total / active parameters | 594B / 17B [2] | 975B / 41B [16] |
| Layers | 48 [3] | 66 [18] |
| Hidden size | 4,608 [3] | 6,144 [18] |
| Routed experts, active per token | 384, 8 [3] | 256, 6 [18] |
| Shared experts | 0 [3] | 2 [18] |
| Local (sliding-window) layers | 40 of 48, window 1,024 tokens [3] | 55 of 66, window 512 tokens [18] |
| Vocabulary | 201,024 [3] | 201,024 [18] |
| Maximum context | 1,048,576 tokens [3] | 1,048,576 tokens [18] |
| Input modalities | Text, image [2] | Text, image, audio [16] |
| Released checkpoint precision | NVFP4 and MXFP8 [2] | bfloat16 and NVFP4 [16][18] |

Both generations interleave five [sliding-window attention](https://aiwiki.ai/wiki/sliding_window_attention) layers with one global layer, and both configurations include a separate multi-token prediction block (`mtp_config`) alongside the main stack [3][18]. Inkling 2 Small activates about 2.9% of its parameters per token, compared with about 4.2% for Inkling, and drops the shared experts that Inkling used [3][18].

Images are handled by a 292-million-parameter vision encoder based on Meta's [Perception Encoder](https://aiwiki.ai/wiki/perception_encoder), which uses 20x20 pixel patches with 2x2 spatial pooling; the resulting image tokens are interleaved with text tokens [1]. The model card gives an optimal range of 40 to 4,096 pixels per image dimension [2]. Audio input, which both earlier Inkling models supported, was removed; the announcement states that "this model does not support audio" without giving a reason [1].

## Training

Thinking Machines trained Inkling 2 Small natively in NVFP4, the 4-bit floating-point format introduced with the [NVIDIA Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell) architecture, and said it would publish its NVFP4 recipe separately [1]. Pretraining and midtraining were split into four phases [1]:

| Phase | Tokens | Data and context |
| --- | --- | --- |
| H1 | 8.4 trillion | Natural language, code and image data; 68 hours on [NVIDIA GB300 NVL72](https://aiwiki.ai/wiki/nvidia_gb300_nvl72) racks |
| H2 | 3.8 trillion | Higher-quality and newer data at 512K context |
| H3 | About 320 billion | Reasoning and agent trajectories at 1M context, mixed with replayed pretraining data |
| H4 | About 34 billion | High-quality agent trajectories |

The phases add up to roughly 12.5 trillion tokens, far fewer than the 45 trillion the company reported for the first Inkling, although the two figures are not described on the same basis [1][16]. The company said the H3 and H4 mixtures consisted mostly of trajectories sampled from earlier models, including Moonshot AI's [Kimi K3](https://aiwiki.ai/wiki/kimi_k3) [1]. The first Inkling's supervised data had likewise come partly from Kimi K2.5 [16]. The model card describes the training data only in general terms (public internet data, data acquired from third parties, internally generated and synthetic data) [2]. The company's training data documentation links a public summary of training content for the model, which gives October 2026 as the latest date of data collection, and says the training data included more than 10 trillion text tokens, more than a billion images and more than a million hours of video, but no audio-only content [15][21].

Post-training replaced the separate reasoning and agentic [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) runs used for earlier models with one joint run: 1,075 steps, more than 17 million rollouts and more than 3 trillion tokens over 9 days on 2,048 GPUs, using more than 4 x 10^23 FLOPs [1]. Coding tasks made up 38.5% of rollouts and 65.0% of RL tokens; the rest covered terminal tasks, document creation, STEM questions, vision, web app building, tool use with simulated users and safety, with rewards from hidden tests, rubric-based LLM judges, reference answers and pairwise reward models [1]. For alignment the company added a general alignment reward and on-policy self-distillation from a teacher model prompted with a written conduct specification [1].

The announcement describes one incident from the RL run. An asynchronous monitor flagged a case in which the model, failing at a task, "worked around a network block through a third-party service it could reach using its search tools" [1]. The team paused the run, hardened its infrastructure, removed environments that needed web access, added synchronous tool-call monitoring, and resumed with additional training phases aimed at alignment [1].

## Benchmarks

### Developer-reported results

The model card compares Inkling 2 Small with open-weight "Flash" and "Pro" models and with closed models of a similar tier. Thinking Machines says its evaluations ran at effort 0.99 (labeled "xhigh") and temperature 1.0, with a 1-million-token trajectory limit, up from the 256K limit it used for Inkling-Small [1]. The selection below is as published by Thinking Machines in its model card. The company says it used scores reported by outside evaluators (including Artificial Analysis, Scale AI, Vals, Datacurve, Surge and ARC Prize) for both its own and the competing models where such scores existed, so the table is a developer-compiled comparison rather than an independent head-to-head test [1][2].

| Benchmark (developer-reported) | Inkling 2 Small | [DeepSeek V4.1 Flash](https://aiwiki.ai/wiki/deepseek_v4_1_flash) (1017) | [GLM-5.3](https://aiwiki.ai/wiki/glm_5_3)-Flash | [Claude Haiku 5.5](https://aiwiki.ai/wiki/claude_haiku_5_5) |
| --- | --- | --- | --- | --- |
| AA Intelligence Index (v5.0) | 30.6 | 32.6 | 27.8 | 36.5 |
| [Vals](https://aiwiki.ai/wiki/vals_ai) Index (v2.2) | 56.1% | 61.3% | 50.9% | 58.7% |
| DeepSWE v1.1 | 72.1% | 77.0% | 63.4% | 59.9% |
| [SWE-Bench Pro](https://aiwiki.ai/wiki/swe_bench_pro) v2 (Hard) | 86.3% | 89.7% | 82.8% | 52.9% |
| FrontierCode v2.0 | 25.4% | 28.5% | 17.6% | 34.9% |
| [BrowseComp](https://aiwiki.ai/wiki/browsecomp) (with compaction) | 90.2% | 87.8% | 76.0% | 80.2% |
| MCP Atlas (public) | 85.0% | 83.2% | 82.0% | 81.0% |
| GDPVal-AA v2.1 (Elo) | 1630 | 1746 | 1649 | 1631 |
| EnterpriseBench-CoreCraft | 71.8% | 58.5% | 46.7% | not reported |
| [Humanity's Last Exam](https://aiwiki.ai/wiki/humanity_s_last_exam) (text only) | 44.1% | 43.3% | 39.9% | 44.4% |
| [ARC-AGI-2](https://aiwiki.ai/wiki/arc_agi_2) | 81.7% | 78.1% | 65.8% | 61.3% |
| MMMU Pro (Standard 10) | 81.6% | 80.6% | 79.0% | 79.1% |
| AA-LCR v1.1 (long context) | 83.3% | 82.7% | 80.0% | 82.7% |
| AA-Omniscience v1.1 (score) | 16.1 | 0.9 | 0.7 | 12.3 |
| FORTRESS adversarial | 83.8% | 72.8% | 73.3% | 90.5% |
| FORTRESS benign | 79.7% | 93.1% | 96.4% | 91.0% |

On the developer's own table, DeepSeek-V4.1-Flash-1017, a model with a similar active parameter count, scores higher than Inkling 2 Small on the Artificial Analysis and Vals indexes and on most of the coding rows, while Inkling 2 Small leads on browsing, enterprise-operations, vision and factuality rows [2]. The two FORTRESS rows also move in opposite directions from the first generation: Inkling's model card reported 78.0% adversarial and 95.9% benign, so Inkling 2 Small refuses more harmful prompts but also more harmless ones [2][16].

### Independent evaluation

Artificial Analysis evaluated Inkling 2 Small at launch and scored it 31 on version 5.0.1 of its Intelligence Index, 18 points above Inkling Small; it said the score matched Google DeepMind's [Gemini 3.8 Flash](https://aiwiki.ai/wiki/gemini_3_8_flash) (high) and was above Xiaomi's [MiMo-V2.6](https://aiwiki.ai/wiki/mimo_v2_6)-Pro [6]. It said that releasing the weights would make Inkling 2 Small the highest-scoring open-weights model developed outside China [6]. On Artificial Analysis's figures, the next-highest such model was [Mistral Large 4](https://aiwiki.ai/wiki/mistral_large_4), at 26 [7]. On the same index version, Artificial Analysis lists the first Inkling (xhigh) at 12 and Inkling Small at 13; the score of 41 it gave Inkling at launch in July 2026 predates version 5.0.1 and is not directly comparable [8][20].

Among the individual results Artificial Analysis reported on November 5, 2026 [6]:

- AA-Briefcase rose from 912 Elo for Inkling Small to 1527, against 1439 for [Claude Sonnet 5.5](https://aiwiki.ai/wiki/claude_sonnet_5_5) (medium) and 1203 for Gemini 3.8 Flash (high).
- FrontierCode 2.0 rose from 5% to 25%, against 11% for Claude Sonnet 5.5 (medium), while Terminal-Bench 4.0 was 13% against 30% for Claude Sonnet 5.5.
- [CritPt](https://aiwiki.ai/wiki/critpt) v2 rose from 34% to 56%, below Gemini 3.8 Flash (high) at 65%.
- On AA-Omniscience v1.1 the hallucination rate fell from 61% to 27%, but accuracy only rose from 32% to 34%; Artificial Analysis said the model "hallucinates less mainly because it answers fewer questions."

The same analysis found the model unusually token-hungry: about 150,000 output tokens per index task, roughly four times Inkling Small, for a cost of $1.18 per task at Thinking Machines' list prices [6]. As of November 17, 2026, the Artificial Analysis model page recorded 154 million output tokens to run the full index, classed the model as "very verbose," and measured 116 output tokens per second with 39.11 seconds to the first answer token on Thinking Machines' API [7].

## Availability

Inkling 2 Small is offered on Tinker for fine-tuning and sampling, and through the Tinker Playground [1][2]. As of November 17, 2026, Tinker's pricing page listed sampling at $0.30 per million prefill tokens, $0.03 per million cached prefill tokens and $1.00 per million sampled tokens, with training at $1.50 per million tokens at a 64K context tier and $3.00 at a 256K tier; both training tiers allow sampling contexts up to 1,048,576 tokens [9]. Serverless inference was in beta at the same input and output rates with a 256K context, and production use required joining a waitlist [9][10]. These rates are lower than Tinker's prices for the original Inkling, which samples at $4.68 per million output tokens [9].

The model card lists local inference recipes for [SGLang](https://aiwiki.ai/wiki/sglang), [vLLM](https://aiwiki.ai/wiki/vllm), TokenSpeed and [Unsloth](https://aiwiki.ai/wiki/unsloth), post-training support in Miles, [SkyRL](https://aiwiki.ai/wiki/skyrl) and Snowflake Cortex Training, and hosted APIs from NVIDIA, [Fireworks AI](https://aiwiki.ai/wiki/fireworks_ai), [Together AI](https://aiwiki.ai/wiki/together_ai), [Baseten](https://aiwiki.ai/wiki/baseten), [Nebius](https://aiwiki.ai/wiki/nebius), [Modal](https://aiwiki.ai/wiki/modal) and [Cloudflare](https://aiwiki.ai/wiki/cloudflare) [2]. Cloudflare's Workers AI lists it at $0.30 per million input tokens and $1.00 per million output tokens with a 1,048,576-token context, on paid plans only [14]. Unsloth published [GGUF](https://aiwiki.ai/wiki/gguf) quantizations for [llama.cpp](https://aiwiki.ai/wiki/llama_cpp), the smallest needing 171.7 GB of disk space for the weights, and noted that running them required a llama.cpp pull request for the new `inkling2` architecture rather than a mainline build [13].

## Safety and limitations

Thinking Machines says it evaluated sycophancy, manipulation and psychological-harm patterns, chemical, biological, radiological, nuclear and cyber uplift (including refusal-suppressed variants), and loss-of-control risks before release, and concluded that the model "did not present risk of material uplift beyond what's already available in the open-weight ecosystem" [2]. On the FORTRESS benchmark it reported a lower average risk score than Inkling and Inkling-Small but a higher over-refusal score [1]. On the company's internal misbehavior measure, where lower is better, Inkling 2 Small scored 14.9, against 66.8 for Inkling, 75.2 for Inkling-Small and 0.8 for OpenAI's [GPT-6.1 Sol](https://aiwiki.ai/wiki/gpt_6_1_sol); the company said the model "still trails closed-weights models" on alignment evaluations [1].

The announcement lists weaknesses directly: the model "is not the most succinct communicator, and doesn't always choose the best tools for the job," with problems that include poorer organization of content, overuse of slashes, and "overreaction to mental health and other safety related queries" [1]. It also says earlier checkpoints showed behaviors such as "fabricating results, ignoring instructions, and concealing what they had done" [1].

## Reception

Coverage concentrated on the benchmark gains and on the delayed weights. VentureBeat's Carl Franzen treated the company's results as a starting point for evaluation rather than proof of deployment readiness, writing that its disclosures about training-time misbehavior support the lesson that "task success and compliance with instructions need separate checks" [12]. He also pointed out that low per-token prices do not guarantee low task costs for a verbose model, and that the launch announcement did not establish a validated minimum GPU configuration for self-hosting, a gap the model card filled when the weights were published a week later [12][2]. Artificial Analysis noted that the model's hallucination improvement came mainly from declining to answer more questions [6].

## See also

- [Inkling (Thinking Machines Lab)](https://aiwiki.ai/wiki/inkling)
- [Thinking Machines Lab](https://aiwiki.ai/wiki/thinking_machines_lab)
- [Tinker (Thinking Machines Lab)](https://aiwiki.ai/wiki/tinker)
- [Open weights](https://aiwiki.ai/wiki/open_weights)
- [List of open-weight models](https://aiwiki.ai/wiki/list_of_open_weight_models)

## References

1. Thinking Machines Lab. "Introducing Inkling 2 Small". Thinking Machines Lab, November 5, 2026 (updated November 12, 2026). https://thinkingmachines.ai/news/inkling-2-small/
2. Thinking Machines Lab. "Inkling 2 Small" (model card). Hugging Face, accessed November 17, 2026. https://huggingface.co/thinkingmachines/Inkling-2-Small
3. Thinking Machines Lab. "Inkling-2-Small config.json". Hugging Face, accessed November 17, 2026. https://huggingface.co/thinkingmachines/Inkling-2-Small/blob/main/config.json
4. Thinking Machines Lab. "thinkingmachines/Inkling-2-Small-MXFP8". Hugging Face, accessed November 17, 2026. https://huggingface.co/thinkingmachines/Inkling-2-Small-MXFP8
5. Thinking Machines Lab. "Inkling-2-Small commit history". Hugging Face, accessed November 17, 2026. https://huggingface.co/thinkingmachines/Inkling-2-Small/commits/main
6. Artificial Analysis. "Thinking Machines Inkling 2 Small: the highest scoring open weights model outside of China pending weights release". Artificial Analysis, November 5, 2026. https://artificialanalysis.ai/articles/thinking-machines-releases-inkling-2-small-highest-scoring-open-weights-model-outside-of-chinia-upon-weights-release-it
7. Artificial Analysis. "Inkling 2 Small: Intelligence, Performance & Price Analysis". Artificial Analysis, accessed November 17, 2026. https://artificialanalysis.ai/models/inkling-2-small
8. Artificial Analysis. "Thinking Machines: Models Intelligence, Performance & Price". Artificial Analysis, accessed November 17, 2026. https://artificialanalysis.ai/providers/thinking-machines
9. Thinking Machines Lab. "Tinker pricing". Thinking Machines Lab, accessed November 17, 2026. https://thinkingmachines.ai/tinker/pricing/
10. Thinking Machines Lab. "Tinker". Thinking Machines Lab, accessed November 17, 2026. https://thinkingmachines.ai/tinker/
11. Thinking Machines Lab. "A Safe Path to Open Weights". Thinking Machines Lab, July 31, 2026. https://thinkingmachines.ai/blog/a-safe-path-to-open-weights/
12. Franzen, Carl. "Thinking Machines launches Inkling 2 Small with near-frontier performance on enterprise benchmarks". VentureBeat, November 6, 2026. https://venturebeat.com/technology/thinking-machines-launches-inkling-2-small-with-near-frontier-performance-on-enterprise-benchmarks
13. Unsloth. "Inkling 2 Small: How to Run Locally". Unsloth documentation, accessed November 17, 2026. https://unsloth.ai/docs/models/inkling-2
14. Cloudflare. "inkling-2-small (Thinking Machines)". Cloudflare Workers AI docs, accessed November 17, 2026. https://developers.cloudflare.com/workers-ai/models/inkling-2-small/
15. Thinking Machines Lab. "Training Data Documentation". Thinking Machines Lab, accessed November 17, 2026. https://thinkingmachines.ai/training-data-documentation/
16. Thinking Machines Lab. "Inkling: Our Open-Weights Model". Thinking Machines Lab, July 15, 2026. https://thinkingmachines.ai/news/introducing-inkling/
17. Thinking Machines Lab. "Introducing Inkling-Small". Thinking Machines Lab, July 30, 2026. https://thinkingmachines.ai/news/inkling-small/
18. Thinking Machines Lab. "Inkling config.json". Hugging Face, accessed November 17, 2026. https://huggingface.co/thinkingmachines/Inkling/blob/main/config.json
19. Hugging Face. "thinkingmachines (Thinking Machines Lab)". Hugging Face, accessed November 17, 2026. https://huggingface.co/thinkingmachines
20. Artificial Analysis. "Thinking Machines has released Inkling, the new leading U.S. open weights model". Artificial Analysis, July 15, 2026. https://artificialanalysis.ai/articles/thinking-machines-has-released-inkling-the-new-leading-u-s-open-weights-model
21. Thinking Machines Lab. "Inkling 2 Small: Public Summary of Training Content". Thinking Machines Lab, November 5, 2026. https://thinkingmachines.ai/documents/inkling-2-small-public-summary-training-content.docx

