Inkling

Inkling is Thinking Machines Lab's open-weight multimodal mixture-of-experts model, a broad foundation model for reasoning, coding, agentic tool use, and fine-tuning.

start building →
compare models

input · usd / 1m tok

$1.87

output · usd / 1m tok

$4.68

context window

65.54ktokens

modalities · in → out

takes text and image in, gives text out

pricing

base api ratesusd per 1m tokens

input
$1.87
output
$4.68
cache read
$0.374

providers

Thinking Machines hosts Inkling. Impossibl routes your request intelligently and handles failover automatically when a provider is down.

  • Thinking Machines

specifications

takes text and image in, gives text outthinkingmachines/inkling

  • creatorThinking Machines
  • model typeText generation
  • context window65,536 tokens
  • maximum outputnot published
  • input

    takes text and image in, gives text out

  • outputtext
  • tool callingnot verified
  • releasedjul 15, 2026
  • endpoint/v1/chat/completions

unknown means the public catalog does not establish that fact. a model's documented capabilities can differ from what a particular api supports. see the model documentation for integration details.

try it · your prompt, this model, live

# responses API
import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["IMPOSSIBL_API_KEY"],
    base_url="https://api.impossibl.com/v1",
)

resp = client.responses.create(
    model="thinkingmachines/inkling",
    input="",
)
print(resp.output_text)

edit the prompt, then run

benchmarks

intelligence index

25.5#44 of 90 models on the gateway

  1. coding index52.1#40 of 60
  2. agentic index24.3#33 of 49
  3. omniscience2.0#32 of 86
  4. gpqa87.2%#43 of 88
  5. humanity’s last exam31.9%#42 of 89
  6. terminal-bench v2.155.1%#40 of 57
  1. GPQA87.2%
  2. AA-LCR77.3%
  3. MMMU-Pro73.5%
  4. Terminal-Bench v2.155.1%
  5. SciCode47.0%
  6. Humanity’s Last Exam31.9%
  7. Tau Banking29.1%
  8. AnalystAgent23.8%
  9. GDP.pdf All-pass12.8%
  10. Multilingual LCR12.2%
  11. CritPt5.4%
  12. AutomationBench Partial5.0%
  13. Terminal-Bench v4.01.0%
  • Intelligence Index25.54
  • Omniscience Index2.00
  • GDPval Elo1165.19
  • coding index (API)52.1
  • agentic index (API)24.3
  • intelligence index (API)25.5

view source ↗all source data and breakdowns (json) ↗

each configuration is listed separately; these evaluations describe the tested model configuration, and performance and costs can differ on impossibl. unmeasured scores are omitted and api values keep their source units. last checked 2026-09-11.

compare Inkling with other models

frontier labs and open weights, best-ranked firstusd / 1m tok · in / out

impossibl models --all →

frequently asked questions

What is Inkling?

Inkling is Thinking Machines Lab's open-weight multimodal mixture-of-experts model, a broad foundation model for reasoning, coding, agentic tool use, and fine-tuning. It processes text, images, audio, and video natively with an encoder-free design and offers controllable thinking effort to trade tokens for performance.

How much does Inkling cost?

Inkling costs $1.87 per million input tokens and $4.68 per million output tokens through the Impossibl API. Cache reads cost $0.374 per million tokens. All rates are in USD.

view the pricing table ↑

What is the context length of Inkling?

Inkling has a 65,536-token context window through the Impossibl API.

What inputs and outputs does Inkling support?

Inkling accepts text and image as input. Inkling returns text.

How do I use Inkling through an API?

Inkling is available through the Impossibl API. Use thinkingmachines/inkling as the model identifier with /v1/chat/completions. The quickstart explains how to create an API key and send your first request.

follow the api quickstart →

When was Inkling released?

Thinking Machines released Inkling on July 15, 2026.

this page describes Inkling as listed by the impossibl public api. prices and availability come from that catalog, refreshed here every five minutes. this is an impossibl offering, not a comparison of independent hosting-provider quotes. model descriptions and release dates, where available, use the creator sources linked under the answers above.

view the source catalog ↗model as markdown ↗report a correction →