Inkling 2 Small explained: 594B MoE, Apache weights pending

By LLMTest Lab · Nov 7, 2026 · 3 min read open-weightsmoenew-models
On this page

On this page

  1. What is new versus the first Inkling
  2. Vendor-reported results
  3. How it compares on paper
  4. Caveats and who should care

Not tested by us yet. Inkling 2 Small is served only through Thinking Machines' own Tinker platform, and OpenRouter lists the older Inkling Small but not this model, so our harness cannot call it. Everything below comes from the announcement and the model card. We'll update this post when an open endpoint or the weights appear and we can run it on our own prompts.

Inkling 2 Small is a 594B-parameter mixture-of-experts model from Thinking Machines Lab with 17B parameters active per token, accepting text and images. It is aimed at agentic coding and long-running work, and the company promises full Apache 2.0 weights the week after external safety testing. Until then it is a hosted-only model, with no published per-token price.

Inkling 2 Small
Maker Thinking Machines Lab
Announced 2026-11-05
Architecture 48-layer sparse MoE, 8 of 384 experts routed per token
Parameters 594B total, 17B active
Input Text and images (the announcement says no audio)
Context Up to 1M tokens per the model card
Access Tinker (inference and fine-tuning); other platforms "in the coming weeks"
License today Tinker Terms of Use; Apache 2.0 weights promised
Price Not published on the announcement page

What is new versus the first Inkling

The company says the model "far surpasses" earlier Inkling models. Its internal misbehavior rate is reported at 14.9, against 66.8 for the original Inkling, and it reached Inkling Small's final performance on a customization task in 75 training steps. These are vendor-reported figures. Pretraining used 8.4T tokens natively in NVFP4, followed by longer-context phases up to 1M tokens, then a 1,075-step reinforcement-learning run on 2,048 GPUs.

For context on the architecture, our explainer on how mixture-of-experts models work covers why 17B active parameters can sit inside a 594B model: you pay compute for the active slice, but you still need memory for all of it.

Vendor-reported results

All numbers here are Thinking Machines' own, run by them or by third-party indices they cite. We have not reproduced any of them.

Benchmark Reported score
Artificial Analysis Intelligence Index v5.0 30.6
SWE-Bench Verified 82.2%
Terminal Bench 2.1 (Pi harness) 85.1%
BrowseComp (with context management) 90.2%
GPQA Diamond 93.0%
MMMU Pro 81.6%
AA-Omniscience non-hallucination rate 73.3%

The model card itself notes that Inkling 2 Small is not top in most rows of its comparison table. It marks the highest score only for BrowseComp, Finance Agent v2.1 (67.5%) and the non-hallucination rate. In one kernel-optimization task the announcement shows Claude Opus 5.5 reaching an 18.80x speedup against 6.02x for Inkling 2 Small. Read the table as "competitive", not "best".

How it compares on paper

Prices below are from OpenRouter's public model list on 2026-11-07, per million tokens.

Model Active params Context Input / output Weights
Inkling 2 Small 17B (594B total) 1M not published Promised, Apache 2.0
Kimi K3.1 Pro not listed there 1,048,576 $2.00 / $10.00 Open
Muse Spark Open not listed there 1,048,576 $1.25 / $4.25 Open
Inkling Small (previous gen) not listed there 524,288 $0.45 / $1.20 No

The previous-generation Inkling Small is the cheap reference point on OpenRouter, but it is a different model, so do not assume 2 Small will price the same.

Caveats and who should care

The company lists its own limits: it is less succinct than some models, less tuned for chat, sometimes picks poor tools in codebases with many skills, and over-disclaims on some mental-health and safety questions. It also says it still trails GPT-6.1 Sol on alignment evaluations. The model card advises against medical, legal or safety-critical use without domain validation.

Who should watch it: teams that want a large open-weight agent model they can fine-tune and self-host, and anyone already on Tinker. Who should wait: anyone who needs a price, an OpenRouter-style endpoint or downloadable weights today, since none exist yet. If you are choosing among open models right now, our roundup of the best open-source LLMs for on-prem use in 2026 is the place to start.

Related articles

Best open-source LLMs for on-prem in 2026: what actually runs where
Six open-source LLMs ranked for on-prem deployment in 2026: hardware minimums, real license terms, and the performance tier you get at each level.
What is MoE? The sparse expert trick behind DeepSeek and Mixtral
Mixture of Experts models run only a fraction of their parameters per token. Here's why DeepSeek and Mixtral are cheap, and when MoE gets expensive.