Inkling Small Guide: 12B Active, and Often Ahead of Inkling

The usual assumption about a "small" model is that you trade quality for price. Inkling Small breaks that assumption on most of the benchmarks its vendor publishes, and the interesting question is not whether it is good but what exactly it gave up to get there.

Thinking Machines Lab released it on 30 July 2026, two weeks after Inkling, describing it as "an efficient open-weights model that achieves comparable performance to Inkling at a quarter of its size." On reasoning and agentic coding the vendor's numbers say it does better than comparable. On factual recall they say it does markedly worse. Both halves matter.

A quarter of the size, mostly higher scores

Inkling Small is a mixture-of-experts transformer with 276B total parameters and 12B active per token, against Inkling's 975B and 41B. It keeps the same encoder-free design, taking text, images and audio directly, with variable thinking effort and a stated context of up to 1M tokens. The weights are on Hugging Face under Apache 2.0, in BF16 and NVFP4.

The vendor's side-by-side table, at maximum effort, puts Small ahead on most of the work developers care about:

• SWE-bench Verified — 80.2% against Inkling's 77.6%
• SWE-bench Pro (public) — 55.9% against 54.3%
• Terminal Bench 2.1 — 64.7% against 63.8%
• Toolathlon Verified — 54.4% against 45.5%
• GPQA Diamond — 89.5% against 87.2%
• ARC-AGI-2 — 40.1% against 36.5%
• IFBench — 82.2% against 79.8%

These are vendor-run figures. Treat the direction as more reliable than the decimals: a two-point gap on a vendor harness can vanish on yours.

Where the larger model still wins

The same table shows what Small traded away, and it is concentrated in one place: knowledge.

• SimpleQA Verified — 20.6% against Inkling's 43.9%, less than half
• AA Omniscience index — -9.0 against 2.1
• Tau 3 Banking — 15.5% against 23.7%
• FORTRESS adversarial — 71.6% against 78.0%, so it refuses adversarial prompts less often

On the aggregate AA Index v4.1 the two are nearly level, 40.0 for Small and 41.0 for Inkling. The vendor puts it plainly: Inkling "maintains an advantage on knowledge coverage and factuality."

That has a concrete consequence. The Omniscience index penalises wrong answers and treats abstaining as neutral, so a negative score means the model gives more wrong answers than right ones instead of saying it does not know. If you use Small for open-domain questions without retrieval, expect fabricated specifics. If you feed it the material, whether a repository, a document set or search results, the weakness largely stops mattering, because the job becomes reasoning over context rather than recall.

Preview numbers versus release numbers

Inkling Small was first shown as a preview inside the Inkling launch post, with a smaller set of scores. Several moved before release: Terminal Bench 2.1 went from 52.7% in the preview to 64.7% at release, SWE-bench Verified from 77.4% to 80.2%, and HLE text-only from 29.6% to 31.6%.

That is a useful reminder about previews in general. A twelve-point move on an agentic benchmark between preview and release is large, and it cuts both ways. If you evaluated the preview and dismissed it, re-run your tests on the released checkpoint. If you see a preview score quoted somewhere, check whether a release figure has replaced it.

What 12B active changes for cost

Active parameters set compute per token, and 12B is less than a third of Inkling's 41B. That shows directly in price. On OpenRouter at the time of writing, Inkling Small lists at $0.45 per million input tokens and $1.20 per million output, with cached input at $0.10; Thinking Machines' own post quotes the same $1.20 output figure against $4.05 for Inkling. On output that is more than a three-fold gap.

For an agent loop that generates a lot of tokens, that multiplier lands on every turn. Active vs total parameters covers why the active figure, not the headline size, predicts serving cost and latency.

Self-hosting is more tractable too. 276B total still needs several GPUs at full precision. Hugging Face lists dozens of community quantisations for both Inkling models, including GGUF builds for llama.cpp, and at under a third of the total size, Small's quantisations fit on far smaller machines than the 975B model's.

Variants and limits worth knowing

• inkling-small:free on OpenRouter is the same model on a rate-limited free endpoint. Use it to evaluate, not to serve users.
• Context varies by provider. The vendor states up to 1M tokens; the paid OpenRouter listing exposes 524,288 at the time of writing.
• Audio length. The Small model card suggests audio clips ideally under 2 minutes, where Inkling's card suggests under 20. If you process long recordings, test that limit before committing.

When to pick Small over Inkling

A short decision rule:

• Pick Inkling Small for coding agents, tool use, reasoning over supplied documents, grading and synthetic data work, and anything high-volume. It is cheaper and, on the vendor's numbers, usually stronger.
• Pick Inkling when answers depend on what the model knows rather than what you give it, when you need stronger refusal behaviour, or when you process long audio.
• Run both behind a router if your traffic mixes the two: send retrieval-backed and coding work to Small, and closed-book questions to Inkling. Model routing and fallbacks covers the pattern.

Whichever you lean towards, test on your own tasks at the effort level you will actually run. The vendor's numbers are at maximum effort, and the gap between the two models may look different at the setting your budget allows.