T
THINKING MACHINES LAB / thinkingmachines/Inkling-2-Mini

Inkling 2 Mini

Thinking Machines Lab multimodal model released on November 20. Its 26.4 billion parameters reuse the middle layers three times per token. The 3.4 billion physically active parameters correspond to about 5.4 billion in effective computation. It accepts text and images and generates text.

01 / SUMMARY

The essential

ORIENTATIONMultimodal model with reused layers

Primary use stated or inferred cautiously from official documentation.

CONTEXT1M tokens

Maximum input capacity when published by the source.

EXITText; separate output maximum unpublished

Documented maximum generation limit.

ACCESSOpen weights and Tinker sampling

Verified availability channels.

MODEL TYPEMultimodal · Code · Text and reasoning

Primary functional family and verified specialties.

LICENSEApache 2.0

Declared terms for API access, model weights, or self-hosting.

02 / TECHNICAL SHEET

Limits and integration

API IDthinkingmachines/Inkling-2-Mini
Model typeMultimodal · Code · Text and reasoning
Access modelOpen weights
LicenseApache 2.0
DeploymentSelf-hosted
Release20 NOV 2026
Knowledge cutoffNot published
EntranceText · Image
ExitText
Context window1M tokens
maximum outputText; separate output maximum unpublished
ReasoningNot published
Published toolsFeatures
Structured OutingsNot published
Batch processingNot published
Prompt cacheNot published
Fine-tuningLoRA on Tinker
Verified platformsHugging Face · Tinker · Transformers · vLLM · SGLang
BEST SUITED FOR

Coding with tools, visual documents and local assistants. The native mixed NVFP4/MXFP8 checkpoint occupies about 20 GB; at least 32 GB of memory is recommended.

WORTH MONITORING

It can enter repetitive reasoning: limit tokens and agent turns. Minimum effort does not disable all reasoning. The million-token sampling window is not the training limit: Tinker offers 64K or 256K for fine-tuning. Memory and speed depend on the runtime.

03 / CAPABILITIES

What it can do

01

Orientation

Coding with tools, visual documents and local assistants. The native mixed NVFP4/MXFP8 checkpoint occupies about 20 GB; at least 32 GB of memory is recommended.

02

Context

1M tokens · Text; separate output maximum unpublished

03

Tools and integration

Features

04

Access

Open weights and Tinker sampling

Input modalities
TextImage
04 / PRICES

Documented cost

ConceptWorthUnit/condition
Standard inputUSD 0.15 / 1M tokensStandard API rate
Cached input0.015 USD / 1M tokensReading reused prefixes
Cache write or storageNot publishedThe condition varies by provider
Standard output0.50 USD / 1M tokensMay include reasoning tokens
Batch inputNot publishedAsynchronous processing
Batch outputNot publishedAsynchronous processing

Tinker charges USD 0.15 input, USD 0.015 cached input and USD 0.50 output per million tokens. Training costs USD 0.75 at 64K or USD 1.50 at 256K; both options sample up to 1M. Quantized formats are not separate models.

i Prices change and may depend on level, region or context length. Check the source before making a decision.

05 / EVALUATIONS

How to read the results

“

The card reports 48.2% on DeepSWE v1.1 with mini-swe-agent v2.4.6 and 74.9% on ten-option MMMU Pro Standard. These use defined protocols; they do not transfer to other checkpoints and are not Inferama tests.

BenchmarkResultMetricSource
DeepSWE v1.148.2 %mini-swe-agent v2.4.6View source ↗
MMMU Pro Standard 1074.9 %Published scoreView source ↗

Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.

06 / VERSIONS

Chronology

thinkingmachines/Inkling-2-Mini

Version added to Inferama's verified catalog.

07 / SOURCES

Traceability