Inkling 2 Small Model Card

1. General information

Model name

Inkling 2 Small

Thinking Machines Lab, Inc.

Date of release

Nov 5, 2026

License

Apache 2.0

Intended uses

Inkling 2 Small is a general-purpose multimodal model that accepts text and image inputs and generates text outputs. It is intended for use in English and other languages, and across multiple coding languages. The model is designed to be used by developers building AI-powered applications, including agentic and tool-use systems, coding assistants, chatbots, and retrieval-augmented generation systems, and is suitable for general-purpose conversational use, instruction-following, and other natural language and visual tasks. It is released with open weights to support research, fine-tuning and integration into third-party products by downstream developers.

Model Acceptable Use Policy

Languages

English, general multilingual support

2. Model properties

Model type

Multimodal autoregressive transformer

Architecture type

A 48 layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) feed-forward backbone: each token is routed to 8 of 384 experts. Attention is a hybrid of local and global layers. The model is natively multimodal – images are encoded via a vision transformer – with image and text inputs projected into a shared hidden space and processed jointly by the decoder.

Parameters

594B total parameters, 17B active

Numerics support

MXFP8 and NVFP4

Input modalities

  • Text: UTF-8 encoded text
  • Image: Any pixel-based image input. For optimal performance, each image dimension should be between 40px to 4096px.

Inkling 2 Small supports a context window of up to 1M tokens.

Output modalities

Text: UTF-8 encoded text

3. Methods of distribution

The weights are available for download through Hugging Face here and here for MXFP8. Developers can also access the model using the Tinker fine-tuning platform, which provides API access.

The model is also available via API access through third party inference providers.

API access to the model

  • For accessing our model via Tinker, you can get started by referring to the Tinker Cookbook here, and installing the tml-renderers package here.
  • Tinker playground also provides inference-only access to the model here.
  • For accessing our model via third party inference providers, you can refer to the documentation from our inference provider partners.

Running the open weights model on your own system

Required hardware

The model is distributed in two checkpoint formats.

The main NVFP4 checkpoint requires a GPU cluster with at least 500GB of aggregated VRAM. This can be met with either of the following example configurations:

  • 2x NVIDIA B300 GPUs
  • 4x NVIDIA H200 GPUs

The alternative MXFP8 checkpoint requires a GPU cluster with at least 800GB of aggregated VRAM. This can be met with either of the following example configurations:

  • 4x NVIDIA B300 GPUs
  • 8x NVIDIA H200 GPUs

The model was pretrained, midtrained and post-trained in NVFP4, so this is the native precision for the model. We recommend the NVFP4 checkpoint for all use-cases, and we’re confident that it’s the best option. The MXFP8 has a very marginal 1–2 percentage points gain over the NVFP4 checkpoint over evals, but we do not think it is meaningful for any practical use-cases.

Required software

Running the model directly on GPU hardware requires an inference deployment framework–either SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face, along with all of their respective dependency libraries.

4. Training

Data types

Training data includes a broad variety of content types, including text, images, video.

Data provenance

Training data for the model was drawn from publicly available sources, acquired from third-parties, generated internally, or synthetically generated or augmented. Publicly available data includes content from the public internet and publicly accessible repositories.

Curation methodologies

Data curation includes cleaning, processing, and modifying datasets. These processing steps, which vary by data type, may include deduplication and filtering to remove junk or other low-quality data, or to advance safety or other objectives.

5. Evaluations

Open Weights FlashOpen Weights ProClosed Weights Flash
Inkling 2 SmallGLM-5.3-FlashMiMo V2.6 FlashMiniMax-M3.1-FlashDeepSeek-V4.1-Flash-1017Muse Spark OpenGLM 5.3MiMo V2.6 ProMistral Large 4Kimi K3GPT-6 LunaGemini 3.8 FlashClaude Haiku 5.5
Model Info
Params (B)activated / total18/32015/30923/42816/55242/81640/75342/102049/1050104/2800–––
Third Party Index
AA Index (v5.0)27.826.729.132.6 (Second-best reported score in this row)24.931.729.926.1–27.930.836.5 (Best reported score in this row)
Vals Index (v2.2)50.9%54.8%53.8%61.3% (Best reported score in this row)48.6%59.7% (Second-best reported score in this row)57.8%56.5%53.0%52.3%56.0%58.7%
Coding
FrontierCode v2.0stirrup v0.2.117.6%21.6%20.8%28.5% (Second-best reported score in this row)7.6%21.6%24.0%15.0%17.3%25.2%18.7%34.9% (Best reported score in this row)
DeepSWE v1.1mini-swe-agent v2.4.663.4%67.9%73.7%77.0% (Best reported score in this row)60.0%69.0%71.9%61.7%68.5%66.6%73.8% (Second-best reported score in this row)59.9%
Terminal Bench Science v0.1mini-swe-agent v2.4.64.8%4.3%3.3%12.9% (Second-best reported score in this row)1.9%9.5%5.7%4.3%8.1%8.6%10.0%20.0% (Best reported score in this row)
Terminal Bench v4.0mini-swe-agent v2.4.632.8%22.7%25.8%33.8%8.1%41.9% (Best reported score in this row)34.9% (Second-best reported score in this row)24.8%12.6%12.6%19.7%32.8%
Terminal Bench v2.1terminus v284.3%76.4%81.7%78.4%76.4%83.9%87.4% (Second-best reported score in this row)79.5%85.0%73.0%87.6% (Best reported score in this row)79.5%
Vibe Code Bench v1.1openhands (vals)30.8%79.0%73.9%87.7% (Second-best reported score in this row)76.2%78.1%85.2%81.6%85.0%81.7%78.7%90.4% (Best reported score in this row)
SWE-Marathon v1.1Inkling 2 Small in Claude Code38.8%––40.0%–42.5%––48.1% (Second-best reported score in this row)28.1%–50.0% (Best reported score in this row)
SWE-Atlas-QnA v1.1mini-swe-agent v2.4.6–––71.0% (Second-best reported score in this row)50.0%–––65.3%50.8%45.2%62.1%
SWE-Atlas-TW v1.1mini-swe-agent v2.4.6–––61.1%41.1%–––63.3% (Second-best reported score in this row)54.4%71.1% (Best reported score in this row)63.3% (Second-best reported score in this row)
SWE-Atlas-Refactor v1.1mini-swe-agent v2.4.6–––54.3% (Second-best reported score in this row)35.7%–––54.3% (Second-best reported score in this row)47.1%55.7% (Best reported score in this row)52.9%
SWEBench Multilingual88.9% (Second-best reported score in this row)79.9%79.9%82.9%77.5%87.6%82.4%79.7%91.4% (Best reported score in this row)78.2%85.8%83.7%
SWEBench Multimodal v2mini-swe-agent (vals)––––––41.3%–44.2% (Best reported score in this row)37.5%40.8%44.0% (Second-best reported score in this row)
SWEBench Pro v2 (Hard)mini-swe-agent v2.4.682.8%89.7% (Best reported score in this row)84.3%89.7% (Best reported score in this row)82.4%84.3%89.2%83.8%88.2%88.7%58.8%52.9%
SWEBench Pro v2 (Full)mini-swe-agent v2.4.6––96.6%–95.0%95.6%––97.7% (Best reported score in this row)–94.9%87.2%
Sudo L7–––––9.4%–13.9% (Second-best reported score in this row)12.2%–10.6%–
Autoresearch
PostTrainBench v1.2Inkling 2 Small in Claude Code36.8% (Best reported score in this row)––30.5%–34.5% (Second-best reported score in this row)––31.5%–––
PMPP-Hard (AUC, 0–10h)pi v0.83.0–––497.36 (Best reported score in this row)–430.96–478.39453.58–––
BioMysteryBenchterminus v2–69.3%––––––72.6% (Best reported score in this row)61.5%62.2%70.4% (Second-best reported score in this row)
Search and Tools
BrowseCompInkling 2 Small with compaction76.0%78.1%69.5%87.8%68.7%75.1%–84.5%88.4%83.6%89.4% (Second-best reported score in this row)80.2%
DeepSearchQAInkling 2 Small with compaction84.3%86.0%82.7%84.5%80.4%84.7%90.0%84.4%95.0% (Best reported score in this row)83.5%90.1%83.2%
MCP AtlasFull / Public–/82.0%–/62.9%–/80.4%–/83.2%89.2/87.5% (Best reported score in this row)84.2/83.6%–/84.4%–/84.5%82.3/84.2%–/77.8%–/87.5%–/81.0%
Knowledge Work
AA-Briefcase v1.11451148714761570 (Second-best reported score in this row)12791503151213481497133712031574 (Best reported score in this row)
GDPVal-AA v2.11649161414651746 (Best reported score in this row)140116571689 (Second-best reported score in this row)13841543144414451631
JobBench60.7%61.2%67.5%68.5% (Second-best reported score in this row)61.8%61.4%62.0%55.2%54.3%49.0%63.4%39.2%
DAYJOB (Healthcare)1.2%––6.0% (Second-best reported score in this row)–4.4%–3.6%4.0%1.2%2.0%–
Enterprise Operations
AutomationBench-AA v1.0.660.4%52.3%72.5% (Best reported score in this row)66.6% (Second-best reported score in this row)49.7%62.2%58.6%57.8%58.3%53.2%59.9%35.4%
EnterpriseOps-Gym33.2%29.0%48.9% (Best reported score in this row)43.6%43.1%36.4%38.0%44.9%45.3%24.9%45.7%25.1%
EnterpriseBench-CoreCraft46.7%––58.5% (Second-best reported score in this row)–51.8%–47.7%43.1%43.1%58.5% (Second-best reported score in this row)–
Finance Work
Finance Agent v2.160.6%63.3%58.8%62.9%60.7%65.0%64.3%61.7%60.0%56.8%65.8% (Second-best reported score in this row)65.4%
Excel Modeling Benchmark v1.166.3%68.9%67.8%69.2%62.0%69.1%67.4%68.7%71.3% (Second-best reported score in this row)70.0%72.9% (Best reported score in this row)65.5%
Tax Agent Bench62.0%59.9%62.7%74.9% (Best reported score in this row)52.1%73.1% (Second-best reported score in this row)64.9%67.5%68.7%58.9%66.8%62.2%
Apex-Accounting v1.1pass@1 / mean score12.3/53.7%–13.5/60.0%12.5/59.3%13.8/58.0%13.9/59.5%–12.0/56.5%9.5/56.7%10.8/53.9%15.5/58.8% (Best reported score in this row)15.3/62.2% (Second-best reported score in this row)
DAYJOB (Finance)7.3%––12.5% (Second-best reported score in this row)–11.0%–4.8%5.5%6.8%4.5%–
DeepFinance–––––27.5%–21.5%–26.4%–40.0% (Best reported score in this row)
Diligence BenchBest harness48.3%–51.6%50.4%59.0% (Best reported score in this row)49.4%51.8%51.5%50.9%42.1%51.1%40.8%
AA-AnalystAgent25.0%––––38.8%37.5%–38.8%43.8%46.3%53.8% (Best reported score in this row)
Tau 3 Banking v1.0.147.2% (Best reported score in this row)–––––––33.4%–––
ATLAS-Finance v1.1–––––3.0%––3.7% (Second-best reported score in this row)–3.7% (Second-best reported score in this row)–
Legal Work
Harvey LAB v1.110.4%13.3%7.5%14.2%25.0% (Best reported score in this row)14.6%13.8%15.8%12.9%4.2%12.1%4.2%
Legal Research Bench45.2%38.0%34.1%45.2%38.5%49.0% (Best reported score in this row)47.1% (Second-best reported score in this row)35.6%46.2%30.3%38.9%43.8%
Visual Creation
DesignArena Web Apps–––––12711274–1280 (Second-best reported score in this row)11601224–
DesignArena Website1266–1275129613001285130312531330 (Second-best reported score in this row)126512951257
DesignArena Game Dev1270–1307126112541292130912181351 (Second-best reported score in this row)125412981314
DesignArena 3D Design1300–1319134412551304132212391367 (Second-best reported score in this row)119912791329
Code Arena (WebDev)1607163715821665 (Best reported score in this row)–1623162515441656 (Second-best reported score in this row)157915831589
AA-3D937––––782977–982–965994 (Second-best reported score in this row)
Code2Video-AA9751052 (Second-best reported score in this row)–––10341052 (Second-best reported score in this row)–104110308991192 (Best reported score in this row)
BVB67.9%––70.0%–––––70.3% (Second-best reported score in this row)65.9%68.4%
Visual Perception
PerceptionBench49.6%39.1%47.1%47.7%44.1%–38.9%46.5%58.5%56.2%68.7% (Best reported score in this row)60.5%
Humanity’s Sixth Sense26.8%25.4%–30.1%29.6%–25.2%–25.5%21.0%41.6% (Best reported score in this row)–
BabyVisionoriginal / with Python53.4/84.0%16.9/68.4%56.1/91.0%44.9/92.5%44.3/80.2%–19.0/73.8%42.3/71.0%53.4/85.7%55.3/79.6%72.8/92.8% (Best reported score in this row)48.3/93.9%
ScreenSpot Prooriginal / with Python–/85.3%35.4/72.5%48.5/76.8%–––56.6/79.7%42.3/43.4%73.8/86.1%81.1/87.7%65.7/93.7%85.4/88.7% (Best reported score in this row)
GroundingME21.3%–––51.1% (Second-best reported score in this row)––38.9%42.2%–––
OmniDocBench v1.692.3%87.4%92.0%91.9%93.8%–88.4%90.2%95.2% (Best reported score in this row)92.3%89.7%92.6%
SimpleVQAwith search and tools67.1%–––63.6%––57.8%68.5% (Best reported score in this row)–––
Visual Reasoning
GDP.pdf15.4%9.0%20.8%17.2%18.6%11.2%19.2%17.6%22.0%22.8% (Second-best reported score in this row)21.0%20.8%
MMMU ProStandard 1079.0%67.8%79.1%80.6%80.9%–74.1%78.0%80.5%79.7%85.6% (Best reported score in this row)79.1%
Charxiv RQoriginal / with Python80.2/89.4%66.2/88.4%82.5/87.6%77.0/89.6%85.4/89.5%–74.2/86.7%83.2/87.4%84.8/91.3%81.8/87.6%86.2/92.4% (Best reported score in this row)84.4/86.3%
Chartographyoriginal / with Python16.3/78.0%7.5/73.9%12.6%/–17.9/83.0%35.7/65.8%–10.8/69.0%21.2/55.2%26.6/66.6%32.0/66.5%40.9/81.0% (Second-best reported score in this row)46.4/86.2% (Best reported score in this row)
MathVision82.9%51.6%89.1%89.4%87.7%–68.3%88.9%93.1% (Second-best reported score in this row)90.0%93.3% (Best reported score in this row)92.1%
VIALSoriginal / with stirrup v0.2.1–/53.0%––/49.7%–/53.2%–––/58.4%–/42.0%20.9%/– (Second-best reported score in this row)–/44.1%–/52.0%–/60.5%
SAGE v1.189.3% (Second-best reported score in this row)83.3%85.3%–85.8%–84.4%74.2%86.8%83.4%88.2%89.8% (Best reported score in this row)
MuirBench71.6%63.2%69.4%75.7%74.5%–72.6%67.5%77.0%73.9%80.1% (Best reported score in this row)74.0%
ZeroBench (main)original / with Python13.0/32.0%7.0%/–21.0%/–23.0%/– (Second-best reported score in this row)14.0/50.0%–8.0%/–15.0/23.0%23.0/41.0% (Second-best reported score in this row)16.0%/–23.0%/– (Second-best reported score in this row)23.0%/– (Second-best reported score in this row)
STEM Reasoning
CritPt v2Full / Hard–/24.5%–/29.5%51.4/32.5%50.6/25.0%52.6/29.5%57.4/33.5%56.9/34.0%50.3/27.0%–/40.5%65.1/42.5% (Best reported score in this row)65.1/46.5% (Best reported score in this row)64.0/43.5%
HLE-Diamond18.2%13.7%28.1%25.3%26.3%17.9%15.2%17.3%26.9%22.2%33.1% (Best reported score in this row)29.6% (Second-best reported score in this row)
HLE (text only)39.9%35.1%48.0% (Second-best reported score in this row)43.3%43.9%42.3%49.4% (Best reported score in this row)36.4%46.9%38.5%47.8%44.4%
GPQA Diamond91.2%88.7%86.7%92.2%89.0%91.7%89.0%89.7%93.5% (Second-best reported score in this row)90.5%95.3% (Best reported score in this row)89.8%
Apex 202526.0%12.0%28.1%68.8% (Best reported score in this row)32.3%29.2%9.9%45.8%65.6% (Second-best reported score in this row)45.8%57.3%17.2%
HMMT Feb 202693.2%87.1%92.4%97.7% (Best reported score in this row)93.9%95.5%88.6%95.5%97.0% (Second-best reported score in this row)90.9%95.5%50.0%
ArxivMath10/2026–––38.3% (Best reported score in this row)19.2%––18.1%––––
BrokenArxiv10/2026–––49.7% (Second-best reported score in this row)33.3%––30.6%––––
ARC-AGI-191.0%––93.5%89.0%–––94.5%86.7%98.5% (Best reported score in this row)89.0%
ARC-AGI-265.8%––78.1%57.5%–––60.4%59.3%89.2% (Best reported score in this row)61.3%
MysteryMechanism–21.6%–––23.0%15.3%––19.4%36.5% (Best reported score in this row)35.1% (Second-best reported score in this row)
Long Context
AA-LCR v1.180.0%74.3%79.7%82.7%78.3%79.7%86.3% (Second-best reported score in this row)78.0%88.7% (Best reported score in this row)83.3%81.3%82.7%
LongBench v268.2%60.7%60.9%69.0%68.7%68.4%68.0%65.7%69.0%65.0%73.7% (Best reported score in this row)65.3%
Factuality
AA-Omniscience v1.1score0.7-10.8-6.40.925.8 (Best reported score in this row)6.0-4.55.1–-8.323.3 (Second-best reported score in this row)12.3
AA-Omniscience v1.1non-hallucination rate61.7%47.8%6.3%35.9%72.0% (Second-best reported score in this row)61.1%39.8%71.8%–20.7%44.9%65.8%
Instruction Following
ComplexConstraints v1.027.6%22.1%23.2%50.4% (Best reported score in this row)41.3%30.5%32.3%36.5%37.9%27.3%48.4% (Second-best reported score in this row)36.8%
Multilinguality
Global-MMLU-Lite89.0%83.9%78.5%90.4%91.2% (Second-best reported score in this row)89.8%88.0%88.0%90.6%90.2%93.2% (Best reported score in this row)87.8%
MDPBench70.9%54.7%69.2%68.8%82.0% (Second-best reported score in this row)–67.0%64.3%83.6% (Best reported score in this row)71.2%81.5%78.4%
Safety
HoneyBench v0.2 (↓)–––76.4% (Second-best reported score in this row)85.6%76.8%–83.6%78.5%–80.5%–
DistressBench72.7%–––87.7% (Best reported score in this row)72.5%–77.0%73.5%66.7%70.7%71.7%
FORTRESS (adversarial)73.3%64.7%*47.9%72.8%92.4% (Best reported score in this row)71.8%69.9%*76.0%73.4%89.2%79.5%*90.5% (Second-best reported score in this row)
FORTRESS (benign)96.4%97.4%* (Best reported score in this row)96.8%93.1%84.0%96.8%97.1%*91.8%97.3% (Second-best reported score in this row)92.4%94.8%*91.0%
StrongREJECT (none)97.8%97.8%97.4%99.4% (Best reported score in this row)97.1%97.5%99.0%99.0%99.0%99.4% (Best reported score in this row)93.0%97.8%
Cybersecurity
CyberBench v1.167.9%75.4%–––72.1%72.9%–75.2%76.3% (Best reported score in this row)43.8%75.5% (Second-best reported score in this row)
PatchEval74.8%77.0%76.5%80.4% (Second-best reported score in this row)77.8%69.6%80.0%76.1%77.0%76.5%77.8%81.7% (Best reported score in this row)
Best reported score in each row (lowest where lower is better); equal displayed scores tieSecond-best reported score in each row; none when first place is tied

* On public split only.

For ScreenSpot Pro evaluation, images whose long side exceed 4,096 pixels are downsampled to that limit while preserving the aspect ratio; the ground truth answers were rescaled accordingly. Smaller images remain unchanged.

6. Safety

We conducted safety evaluations ahead of release, spanning both everyday human-AI interaction and dangerous-capability testing. Because Inkling 2 Small is multimodal, we paid attention to whether safety behavior held consistently across text and image inputs. We applied mitigations to reduce risks before release.

For everyday interaction, we evaluated sycophancy, harmful manipulation, and psychological-harm patterns like parasocial dependency and validation of delusional reasoning, including through multi-turn, open-ended external red-teaming designed to surface issues that only emerge over longer conversations. We also assessed whether the model refuses genuinely harmful requests without over-refusing benign ones. For CBRN and cyber, we assessed knowledge and procedural uplift through internal evaluations, external testing, and refusal-suppressed variants intended to estimate latent capability with safeguards removed. For loss of control, we evaluated agentic capability, strategic deception, and sabotage potential, benchmarked against public frontier models, and found the model materially below frontier capabilities.

Across all areas, we concluded that Inkling 2 Small did not present risk of material uplift beyond what’s already available in the open-weight ecosystem.

7. Bias, risks and limitations

Inkling 2 Small may exhibit general limitations common to foundation models, including hallucination (generating plausible but factually incorrect or unsupported content), occasional failures to follow instructions precisely, and degraded performance in long multi-turn conversations. As with other large-scale models trained on web-derived and synthetic data, Inkling 2 Small may reflect biases present in its training data, including demographic, cultural, or linguistic biases, and may perform unevenly across languages, dialects, or subject domains that were less represented during training.

Inkling 2 Small’s knowledge is limited to information available as of its training cutoff, and it may not reflect events, developments, or changes that occurred afterward.

We recommend that downstream developers and deployers apply appropriate human oversight and review for outputs used in high-stakes or safety-critical contexts, rather than relying on Inkling 2 Small’s outputs without verification.

  • Conduct their own evaluation of Inkling 2 Small’s performance, safety, and fairness for their specific use case, language, and population prior to deployment, particularly for applications involving vulnerable groups.
  • Implement additional safeguards – such as content filtering, rate limiting, and monitoring – at the application layer, especially for open deployment contexts where Inkling 2 Small’s built-in mitigations may not be sufficient on their own.
  • Avoid deploying Inkling 2 Small in domains such as medical, legal, or safety-critical decision-making without additional fine-tuning, domain-specific validation, and human oversight

Training Data Documentation