Sign in
Inkling logo

Inkling

Thinking Machines Lab · Released Jul 2026

View pricingCompare models
Conditional

Inkling is the strongest reason yet to take open-weight multimodal AI seriously. It delivers competitive reasoning and coding with best-in-class open-weight safety and controllable thinking effort, all under Apache 2.0. But Thinking Machines is honest that it's not the strongest overall — factuality lags (SimpleQA 43.9%) and self-hosting demands serious hardware (2TB VRAM). Conditional: a compelling base for organizations that need to own and fine-tune a multimodal model, not for teams wanting peak off-the-shelf performance.

Is it right for you?

Good for

  • Open-weight (Apache 2.0) multimodal base — full weights downloadable, modifiable, and deployable on your own infrastructure
  • Safety-conscious deployments — best open-weight FORTRESS scores (78.0% adversarial, 95.9% benign refusal)
  • Token-efficient reasoning via controllable thinking effort — uses 1/3 the tokens of Nemotron 3 Ultra for equal Terminal Bench performance
  • Domain-specific fine-tuning on Tinker — proven uplift in financial reasoning with Bridgewater Associates
  • Native multimodal input (text, image, audio) — rare among open-weight models at this capability level

Not good for

  • Pure factuality and knowledge recall tasks — SimpleQA Verified only 43.9%, well behind closed models at 70%+
  • Budget-constrained self-hosting — BF16 checkpoint requires ~2TB aggregate VRAM (8x B300 or 16x H200)
  • Maximum raw coding leaderboard performance — SWE-bench Pro 54.3% and Terminal Bench 63.8% trail closed frontier

How it performs by task

Code generation

Very Good

Solid SWE-bench Verified 77.6%, competitive with open-weight leaders; trails closed frontier on harder repos

Reasoning

Very Good

AIME 2026 97.1% is near-saturated; HLE 29.7% shows limits on hardest reasoning vs. 53%+ from closed leaders

Multimodal understanding

Very Good

Strong vision (MMMU Pro 73.5%, Charxiv 82%) with rare native audio input; limited to text output

Agentic tool use

Very Good

MCP Atlas 74.1% and BrowseComp 77.1% show capable agentic behavior across diverse tool ecosystems

Safety and refusal

Excellent

Best-in-class among open-weight on FORTRESS; calibrated uncertainty instead of hallucination on ambiguous queries

Factuality

Fair

SimpleQA 43.9% is a clear weakness vs. closed models at 70%+; factual reliability needs fine-tuning

Pricing

Input

$1.87 / 1M tokens

Output

$4.68 / 1M tokens

Context

Up to 1M tokens native; Tinker 64K and 256K tiers; rates at 50% limited-time discount (list $3.74 / $9.36)

View full pricing

Benchmarks

BenchmarkScoreSource
AIME 202697.1% Source
GPQA Diamond87.2% Source
HLE (text only)29.7% Source
SWE-bench Verified77.6% Source
Terminal Bench 2.163.8% Source
MCP Atlas74.1% Source
MMMU Pro (Standard 10)73.5% Source
IFBench79.8% Source
SimpleQA Verified43.9% Source

No verdict changes yet

The clock starts day one. Changes land here as our verdict evolves.

Sources

Verification log

  • Pricing· No changes

    Automated agent

  • Pricing· Updated

    Automated agent

    Resolves the 2026-10-05 ambiguity: the stored $1.87/$4.68 are Tinker prefill and sample rates under a 50% limited-time discount (list $3.74/$9.36), identical for the 64K and 256K tiers. pricing_context now says so. Stored rates unchanged.

  • Pricing· Needs attention

    Automated agent

  • Pricing· Needs attention

    Automated agent

    Tinker pricing tiers for Inkling remain ambiguous (carried from 09-26); stored $1.87/$4.68 could not be tied to one tier, so it is left unchanged and no update is written.

  • Pricing· Needs attention

    Automated agent

    The live Tinker page (summarised read) lists four Inkling rows per 1M tokens. Tinker 64K: prefill $3.74, sample $9.36, train $11.22. Tinker 256K: $7.48 / $18.72 / $22.46. Serverless Inference (Beta) Inkling 256K: $1.00 in, $0.17 cached, $4.05 out. Inkling-Small serverless: $0.30 / $1.20. The page also notes a 'limited-time 50% discount'. The stored $1.87/$4.68 equals half the 64K Tinker prefill/sample, but it is unclear whether the listed numbers are before or after the discount, or which row our scalar should track. Stored values are kept; this needs a human read. Also noted, not written: 975B total / 41B active, 256K context rows (stored 'Up to 1M tokens native').

  • Pricing· No changes

    Automated agent

  • Pricing· No changes

    Automated agent

How we evaluate