Inkling-Small: Thinking Machines’ 12B-Active MoE Matches Inkling at 1/4 Size

Thinking Machines Lab released Inkling-Small on July 30, 2026 — full open weights for a 276B MoE with 12B active parameters that matches or beats full Inkling on key reasoning and agentic benchmarks at roughly a quarter of the size.

That is the clean follow-up to our July Inkling launch guide: the “preview sibling” is no longer a footnote. Weights are live on Hugging Face, fine-tuning is on Tinker, and the lab’s own tables show Small ahead on Humanity’s Last Exam (text-only) and SWE-bench Verified while trailing badly on SimpleQA. This explainx.ai post is the practical map — when Small wins, where Inkling still wins, and how the efficiency curves change agent economics.

TL;DR — questions after the Jul 30 post

Why a “Small” that still has 276B parameters

MoE naming confuses people. Active parameters drive most of the FLOPs per token; total parameters are capacity sitting behind routing. Inkling-Small’s claim is not “tiny laptop model” — it is comparable quality at ~12B active versus Inkling’s 41B active, with full weights you can fine-tune.

Thinking Machines frames the performance-compute story with effort sweeps (minimal → xhigh) on Terminal-Bench 2.1, HLE (no tools), and IFBench. Output TFLOPs per sample are estimated as 2 × active parameters × mean generated tokens (including reasoning tokens). The point of the charts: Small’s curve is competitive with other open-weights models in its weight class and more efficient than full Inkling for many operating points.

If you already care about open-weight customization and the lab’s human-judgment manifesto, Small is the deployment-friendly sibling of that bet — same product philosophy, cheaper trajectories.

Training story that explains the weird scoreboard

Inkling-Small started training after Inkling, which let the lab change the pre-training data mix and ML recipe. They post-trained an earlier Inkling-Small (preview) checkpoint partly with on-policy distillation using Inkling as the teacher, then continued agentic coding RL for about two weeks. After those improvements, Small surpassed Inkling on reasoning and agentic coding benches. Inkling keeps the edge on knowledge coverage and factuality.

That is why the scoreboard looks “inverted” if you only remember July 15 preview numbers. Preview Terminal Bench was weaker; the Jul 30 full release shows Small competitive or ahead on several agent rows. Always pin dates when you compare tables.

Benchmark snapshot (effort 0.99, lab table)

Selected rows from Thinking Machines’ full suite (they also cite Artificial Analysis, Scale AI, ARC Prize, ForecastBench, Prophet Arena where applicable). Coding evals use a 256K max-token trajectory limit.

*Harness footnotes matter — see the announcement. External models often use self-reported numbers; Inkling family uses bash-only / internal coding harnesses. Treat cross-vendor rows as orientation, not a contract — same caution as our AI benchmarks guide.

Against open peers in a similar active-parameter band (Qwen3.5 397B-A17B, MiMo V2.5, Minimax M2.7, DeepSeek V4 Flash, Nemotron 3 Ultra), Small lands as a broad generalist: strong coding + reasoning efficiency, competitive multimodality, weaker pure knowledge than the largest open and closed systems.

Multimodality without a bolted-on encoder

Same architecture family as Inkling: encoder-free multimodal MoE. Audio as dMel spectrograms; images as 40×40 patches through a four-layer hMLP; light embedding into the joint token stream. The lab highlights audio intelligence and better Python-assisted visual workflows — crop, zoom, programmatic inspection for small chart text.

That matters for agent harnesses that already treat vision as tool-mediated (screenshot → inspect → act), not only as a single forward pass. Pair it with how we talk about agent harness layers and OpenCode: cheaper active params mean you can afford more tool rounds at the same budget.

Epistemics and safety (inherited recipe)

Calibration training used RL against proper scoring rules on real-world forecasting questions. ForecastBench / Prophet Arena numbers put Small in the same band as Inkling — slightly ahead on some ForecastBench rows in the Jul 19–28, 2026 test window.

Safety post-training matches Inkling’s internal spec (everyday interactions + dual-use), with internal evals and external red-team partners. StrongREJECT is near-parity; FORTRESS adversarial is a bit softer than Inkling (71.6% vs 78.0%) while benign answering stays high (96.9%). If you ship customer-facing agents, run your own refusal/over-refusal suite — public benches are not a threat model.

How to run it (practical paths)

1. Playground / API: Tinker Playground for text, image, and audio chat; Tinker for fine-tuning (limited-time discount called out in the launch post).
2. Weights: thinkingmachines/Inkling-Small (+ NVFP4 variants for Blackwell-class hardware).
3. Self-host: SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face stacks per the model card.
4. Harness: Point OpenCode or your internal agent at a provider endpoint; sweep effort the way you would for full Inkling.

# Example: pull the HF repo id into your usual inference stack
# (exact flags depend on vLLM / SGLang version — check the model card)
huggingface-cli download thinkingmachines/Inkling-Small

For local open-weight workflows, see how to run open-source models with OpenCode and the enterprise open alternatives map.

Decision guide: Small vs Inkling vs other open MoEs

Also situate against DeepSeek Flash traffic economics and Kimi coding specialists: different Pareto frontiers (price, coding specialization, multimodality). Small’s differentiator is Inkling-family breadth + effort dial + Tinker at 12B active.

Honest limitations

• Factuality gap is large. Do not treat Small as a drop-in knowledge base replacement for Inkling.
• Harness footnotes. Terminal Bench and SWE numbers depend on harness choice; contaminated web-search solutions were zeroed for Inkling family.
• Hardware reality. 276B total still needs serious inference infrastructure even if 12B are active — this is not a 7B laptop default.
• Safety is not solved. Competitive StrongREJECT/FORTRESS ≠ your policy.

What “variable thinking effort” changes in agent budgets

Full Inkling already shipped an effort dial; Small keeps the same product idea with cheaper active FLOPs underneath. For coding agents that emit long reasoning traces, that matters more than headline parameter counts. A harness that used to burn 41B-active tokens on every “think hard” step can now spend 12B-active tokens at a comparable quality band on many agentic rows — then reserve full Inkling (or a closed frontier model) for knowledge-heavy subcalls.

A practical pattern we see teams use with OpenCode and similar harnesses:

1. Default the coding agent to Inkling-Small at medium effort for edit/test loops.
2. Escalate to xhigh / 0.99 only when Terminal-Bench-style tasks or hard SWE tickets stall.
3. Route factual lookup and long-tail knowledge to full Inkling or a search-augmented path — Small’s SimpleQA gap is the tell.
4. Fine-tune Small on Tinker for one domain workflow (support macros, internal APIs, style rules) before you fine-tune the 41B-active sibling.

That is the same customization thesis as the manifesto, just with a more deployable base. If your org already compared open MoEs after the American open-weights letter, add Small to the bake-off with an explicit cost column: dollars per solved ticket at fixed harness, not only pass@1.

Audio and vision make the economics sharper. Native speech understanding without a separate ASR stack means voice agents can keep one model in the loop for “hear → reason → tool → speak plan.” Chart and document workflows that call Python for crop/zoom benefit when the model expects that tool pattern — Small’s CharXiv-with-Python lift is the published signal. Pair multimodal sessions with the same sandbox discipline you would use for any tool-using agent: see our notes on sandbox escapes and destructive command guards before you grant shell on production repos.

Related on explainx.ai

• Inkling open weights launch guide (July 15)
• Thinking Machines manifesto — Future Worth Building Is Human
• Open Weights & American AI Leadership letter
• Fable 5 & GPT-5.6 open-source alternatives
• GPT-5.6 vs Claude Fable 5
• OpenCode coding agent guide
• AI benchmarks complete guide
• DeepSeek Flash — 8T tokens/day economics
• What are LLM parameters? Top model sizes

Primary sources

• Introducing Inkling-Small
• Inkling-Small model card
• Hugging Face — thinkingmachines/Inkling-Small
• Inkling product page (both sizes)

Benchmarks, parameter counts, and availability reflect Thinking Machines Lab’s July 30, 2026 announcement and Hugging Face card as checked August 3, 2026. Effort settings, harness variance, and fine-tuning change results — verify current docs before production commitment.