The Model Gap

Thinking Machines Lab

Inkling 2 Small

Thinking Machines released Inkling 2 Small on 2026-11-05: 594B total, 17B active, text and image in, text out, no audio. Tinker serves it at $0.30/$1.00 per million tokens on 64K and 256K windows. The model card's 1M is not served. Weights are closed; Apache 2.0 is promised after external safety testing, with no date. No tracked score is stored. Artificial Analysis displays Intelligence Index 31 (v5.0.1); that composite stays prose.

Inkling 2 Small benchmarks and pricing, every number sourced: no Inkling 2 Small benchmark scores tracked yet, priced at $0.30 per million input tokens and $1.00 per million output.

Inkling 2 Small architecture: Sparse mixture-of-experts, 8 of 384 experts; text and image in, text out; 594B total parameters (17B activated per token); 262K-token context window.

Released
2026-11-05
Availability
Tinker inference and fine-tuning since 2026-11-05. Weights closed. Apache 2.0 planned after external safety testing, no date set. Not a general public API.
License
proprietary
Context window
262K tokensTinker models.json and serverless.json, checked 2026-11-07, serve 64K (thinkingmachines/Inkling-2-Small) and 256K (thinkingmachines/Inkling-2-Small:peft:262144, and serverless thinkingmachines/Inkling-2-Small:peft:262144:sampling-nvfp4). The model card claims up to 1M. No checked route serves 1M. Stored value is the served ceiling.
Knowledge cutoff
Not disclosed
Parameters
594B (17B active)
Architecture
Sparse mixture-of-experts, 8 of 384 experts; text and image in, text out

Inkling 2 Small announced pricing

$0.30 in / $1.00 out per 1M tokens — official pricing source

What Inkling 2 Small costs per job

Inkling 2 Small cost for three reference workloads, computed from its list rates
WorkloadTokens in / outCost
One long chat turn100K / 10K$0.040
A codebase review1,000K / 100K$0.400
A day of agent work10,000K / 1,000K$4.00

Computed from Inkling 2 Small’s list rates above — cache discounts and batch tiers are not applied.

Notes on the record

Thinking Machines Lab released Inkling 2 Small on 2026-11-05. Access today is Tinker, for inference and fine-tuning. The model card's license line is Tinker Terms of Use. The launch post says weights are closed and that Apache 2.0 weights are intended next week, after external safety testing and red-teaming. No repository and no date were published, so this page records proprietary. Artificial Analysis's licenseName is Apache 2.0 on an object that also sets isOpenWeights false; the launch post is what this page follows.

Tinker's models.json, checked 2026-11-07, lists prefill $0.30 and sample $1.00 per million tokens on thinkingmachines/Inkling-2-Small (64K) and thinkingmachines/Inkling-2-Small:peft:262144 (256K). Cached prefill is $0.03. Training is $1.50 at 64K and $3.00 at 256K, and the docs say extended context costs more for training only. Serverless inference, id thinkingmachines/Inkling-2-Small:peft:262144:sampling-nvfp4, is 256K at the same $0.30 and $1.00. The slug on this page is not a Tinker id. No long-context inference multiplier is published.

The model card states 594B total and 17B active parameters, text and image input, text output, and a context window of up to 1M tokens. The launch post says the model does not support audio, and it describes a 292M-parameter vision encoder. Training data includes video; that is not an input. Tinker labels this model Hybrid + Vision. It serves 64K and 256K. This page stores 262,144 and does not store 1,000,000. No knowledge-cutoff date is published.

No score row is added. Artificial Analysis's model page, checked 2026-11-07, displays Intelligence Index 31 under v5.0.1; the embedded value is 30.6286781498958, the launch table labels AA Index (v5.0) at 30.6, and the blog lede says 31. No predecessor index is stored, and v5 is not compared with any pre-v5 reading. Other launch-table cells, including two Terminal-Bench 2.1 harnesses, stay vendor claims. Early checkpoints were tested on OpenCode as Fledge-Alpha. That does not rewrite the 2026-10-01 tracker entry, and nothing here says Inkling 2 Small replaces Inkling or Inkling-Small.

Compare with

GPT-6 CyberGemini 4 CarbonGPT-6.1 TerraKimi K3.1 Proany model…

FAQ

What inputs does Inkling 2 Small accept?

Text and image in, text out. The launch post says the model does not support audio. The model card's training-data section includes video, which is not an input modality. Tinker labels the served model Hybrid + Vision, and labels the older Inkling ids Hybrid + Audio + Vision.

Can I download Inkling 2 Small's weights?

No. The launch post says the weights are closed at release and that full weights under Apache 2.0 are intended next week, after external safety testing and red-teaming. No date and no repository were published as of 2026-11-07. This page therefore records the license as proprietary. Artificial Analysis's licenseName field says Apache 2.0 while isOpenWeights is false; the launch post is the source used here.

What does Inkling 2 Small cost on Tinker?

Prefill is $0.30 and output is $1.00 per million tokens, on both the 64K and 256K tiers, checked 2026-11-07 in Tinker's models.json and serverless.json. Cached prefill is $0.03. Training is $1.50 at 64K and $3.00 at 256K. The docs say the higher training price is training only. No long-context inference step is published. The ids are thinkingmachines/Inkling-2-Small, thinkingmachines/Inkling-2-Small:peft:262144, and the serverless sampler thinkingmachines/Inkling-2-Small:peft:262144:sampling-nvfp4.

Does Inkling 2 Small serve a 1 million token context window?

The model card says up to 1M tokens. Tinker does not serve that. The training catalog lists 64K and 256K, and serverless inference is 256K. This page stores 262,144, the 256K tier in the id thinkingmachines/Inkling-2-Small:peft:262144. Artificial Analysis's model object sets contextWindowTokens to 1024000, which repeats a 1M claim. That figure is not the served window.

Has Inkling 2 Small been independently benchmarked on a board this site tracks?

No tracked score is stored. Artificial Analysis's model page, checked 2026-11-07, displays Intelligence Index 31, from an embedded 30.6286781498958, under Intelligence Index v5.0.1. The launch table labels AA Index (v5.0) at 30.6, and the blog lede and the X post say 31. That composite is not a tracked benchmark, and it is not compared with any pre-v5 absolute. The launch table's other cells stay vendor claims.

Does Inkling 2 Small replace Inkling Small, or is it Fledge Alpha?

No supersession is stated. Tinker's catalog still lists Inkling and Inkling-Small. The launch post says early checkpoints of Inkling 2 Small were tested on OpenCode under the name Fledge-Alpha. That sentence is a cross-reference. It does not identify the model served as Fledge Alpha on 2026-10-01 as this release, and the tracker entry for that date is left as written.

Further reading