Thinking Machines language model

Inkling

Inkling is Thinking Machines' open-weights model and the first in its family: a 975B-parameter mixture-of-experts (41B active) that reasons over text, images and audio with controllable thinking effort. Inkling Small is its smaller, cheaper sibling.

01

price

$1/M

input tokens

$4.05/M

output tokens

$0.17/M

cached input

per 1M tokens. octomind bills the listed provider price, with no markup, on every plan, including the free one (pay as you go from credits).

02

what real work costs

workloadcost
one chat reply 3,000 tokens in, 500 out $0.005
summarise a 100-page report 50,000 tokens in, 1,500 out, one request $0.06
1,000 classification calls 500 tokens in, 20 out, per call $0.58
coding-agent session 2M tokens in across ~60 turns, 60,000 out, no cache $2.24
same session, cached 85% of input read from the prompt cache, the rest written to it $0.83

at the listed rate. token counts are typical sizes, not measurements.

03

specs

provider
Thinking Machines
context window
256k tokens — ≈ 380 pages of prose or 26,000 lines of code
input
text, images
output
text
structured output
yes — JSON schema
reasoning mode
yes
prompt caching
yes — cache hits at $0.17/M
weights
open — Apache 2.0
released
July 15, 2026
status
live
model id
inkling

integration notes

  • thinking has no hard off switch: effort 0.0 only steers the model toward no reasoning; omitted effort defaults to 0.9.
  • reasoning effort and max_tokens are independent, so high effort can truncate output unless you raise the token budget.
  • images must be PNG or JPEG; HTTP and cloud-storage image urls are not fetched, so download the image and send a base64 data URI.

04

benchmarks

as published by the maker; not measured by octomind.

benchmarkscore
SWEBench Verifiedsource77.6%
SWEBench Pro (Public)source54.3%
Terminal Bench 2.1 (Best Harness)source63.8%
AIME 2026source97.1%
BrowseComp (w/ ctx management)source77.1%

05

when to pick it

pick it if

  • you want Apache 2.0 open weights you can later fine-tune on Tinker
  • you send images with text and want to tune thinking effort
  • you run coding or agent harnesses, the work it was trained for

06

where it sits on price

blended price $1.762/M at 3 input tokens per output token: 24 live language models cost less, 19 cost more.

  1. Kimi 2.6 $1.712/M
  2. Kimi 2.7 Code $1.712/M
  3. Inkling $1.762/M
  4. DeepSeek V4 Pro $1.98/M
  5. GPT-6.1 Terra $2/M

07

use Inkling in octomind

one octohub login reaches every model in the catalog. pick this one per run, per session, or as your default.

run it
octomind login
octomind run -m octohub:inkling
switch mid-session
/model octohub:inkling
make it the default — config.toml
[model]
name = "octohub:inkling"

08

compare Inkling

all model comparisons →

09

similar models

nearest live models by listed input price.

view all models →

10

Inkling faq

how much does Inkling cost?

Inkling lists at $1/M input, $4.05/M output, $0.17/M cached input per 1M tokens. a typical coding-agent session (2M tokens in, 60,000 out) comes to $2.24, or $0.83 when 85% of the input is read from the prompt cache. octomind bills the listed provider price, with no markup, on every plan, including the free one (pay as you go from credits).

what is the context window of Inkling?

256k tokens, ≈ 380 pages of prose or 26,000 lines of code.

does Inkling accept images?

yes. it takes text and images as input and returns text.

is Inkling open source?

its weights are open under Apache 2.0, so it can also be self-hosted.

when was Inkling released?

July 15, 2026, by Thinking Machines.

is Inkling free?

no. Inkling bills per token at $1/M input, $4.05/M output, $0.17/M cached input per 1M. octomind's free plan reaches it pay as you go from credits, so you pay only for the tokens you use, at the provider price. credits carry a 3% top-up fee.

how do i use Inkling in octomind?

run octomind login once, then octomind run -m octohub:inkling. inside a running session, /model octohub:inkling switches to it without losing context.

use Inkling in octomind

choose a model when you start, then switch mid-session when the work changes.