Introducing Inkling-Small — Thinking Machines' efficient open-weights MoE reasoning model

TL;DR

Inkling-Small is Thinking Machines’ newly released open-weights sibling to Introducing Inkling — Thinking Machines' open-weights 975B/41B MoE multimodal reasoning model — a 276B-total / 12B-active MoE transformer, ~¼ the total parameters of Inkling, that matches or exceeds its larger counterpart on many reasoning and agentic benchmarks while running considerably cheaper. It inherits Inkling’s encoder-free multimodal recipe (dMel spectrograms for audio, 40×40 hMLP patches for vision), 1M context window, and variable thinking-effort knob. The differentiator vs the July release is process: Inkling-Small started training after Inkling and got an improved pre-training data mix, then post-trained via on-policy distillation from Inkling as teacher, plus two more weeks of agentic coding RL — a live datapoint on the On-Policy Distillation recipe scaling to frontier-class training. Ships with full weights on HuggingFace and Tinker fine-tuning + Playground access.

Key claims

Architecture inherits Inkling wholesale — same MoE recipe, same encoder-free multimodal design, same 1M context, same variable thinking-effort knob — with parameter count scaled down to 276B total / 12B active [§Introduction, §Multimodality].
Trained on NVIDIA GB300 NVL72, same hardware target as Inkling [§Introduction].
Post-training pipeline: on-policy distillation with Inkling as teacher on an earlier Inkling-Small checkpoint, followed by two additional weeks of agentic coding RL — after which Inkling-Small surpasses Inkling on reasoning and agentic coding benchmarks [§Capabilities].
Inkling maintains an advantage over Inkling-Small on knowledge coverage and factuality — the smaller model does not dominate uniformly, only on the reasoning + agentic dimensions where the extra RL was invested [§Capabilities].
HLE (text-only, no tools) 31.6% at effort=0.99, up from Inkling’s 29.7%; test-time-compute curves show Inkling-Small ≥ Inkling at every thinking budget, not just at max effort [§Reasoning and Agentic Tasks].
SWEBench Verified > 80%, exceeding Inkling’s 77.6% [§Reasoning and Agentic Tasks, main table].
Multimodal near-parity: MMMU Pro Standard 10 74.0% (vs Inkling class), Charxiv RQ 77.4/81.3% (without/with Python), Audio MC 54.9%, MMAU 77.0%, VoiceBench 90.1% [§Multimodality, table].
Vision Python-tool augmentation: Inkling-Small can call Python for image cropping, zooming, and programmatic inspection during reasoning — Charxiv RQ jumps from 77.4% (no tool) to 81.3% (with Python) [§Multimodality, table].
Forecasting/calibration: ForecastBench Brier Index (no search) 61.3 ± 0.46, exceeding Inkling’s 60.1 ± 0.54 and matching Grok 4.3 / Gemini 3.1 Pro; trained on real-world forecasting questions with proper-scoring-rule RL [§Epistemics, table].
Safety numbers within a couple of points of Inkling: StrongREJECT 98.4% (vs 98.6%), FORTRESS Adversarial 71.6% (vs 78.0%), FORTRESS Benign 96.9% (vs 95.9%) — same safety post-training recipe [§Safety, table].
Weights released on HuggingFace under thinkingmachines/Inkling-Small; fine-tuning available on Tinker; text/image/audio chat on Tinker Playground [§Introduction, §Try Inkling-Small on Tinker].

Method

Inkling-Small is a process paper rather than a new-architecture paper — the model changes vs Inkling are (a) parameter count and (b) training-recipe refinements enabled by learnings from the larger sibling. Concretely, the pre-training data mix and ML recipe were updated after seeing Inkling’s training dynamics, then an earlier Inkling-Small checkpoint (previously called “Inkling-Small (preview)” and released alongside Inkling in July) was resurrected as the base for a post-training push. That push has two phases: on-policy distillation with Inkling as teacher — the same recipe from the Thinking Machines Tinker-team write-up — then two additional weeks of agentic coding RL specifically to push SWEBench and Terminal-Bench numbers.

The multimodal front end is unchanged: audio is discrete dMel spectrograms (Bai et al. 2024, same as Inkling), images are 40×40 patches processed by a 4-layer hMLP (Touvron et al. 2022), both combined with text tokens through a lightweight embedding layer and processed jointly by the MoE transformer. No frozen vision encoder, no projector — same encoder-free bet as its big sibling and Gemma 4 12B (Gemma 4 12B — encoder-free unified multimodal model with 35M vision embedding (Google Gemma)), Tuna-2 (Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation).

Results

HLE text-only 31.6% at effort=0.99 (Inkling: 29.7%); Inkling-Small’s TTC curves sit above Inkling’s at every thinking budget [§Reasoning and Agentic Tasks].
SWEBench Verified > 80% (Inkling: 77.6%); Terminal Bench 2.1 uses an internal harness (contamination-adjusted); numbers positioned as competitive with same-weight-class open models (MiMo V2.5 15B/310B, Minimax M2.7 10B/230B, DeepSeek V4 Flash 13B/284B) [§Reasoning and Agentic Tasks, main table].
ForecastBench Brier Index (no search) 61.3 ± 0.46 (Inkling 60.1 ± 0.54, Grok 4.3 60.9 ± 0.38, Gemini 3.1 Pro 60.6 ± 0.43); with search 61.5 ± 0.54 [§Epistemics].
Prophet Arena Brier Score 0.1238 ± 0.0086, ahead of Inkling (0.1276) and Kimi K2.6 (0.1265) [§Epistemics].
Multimodal vs specialist omni models: outperforms Nemotron-3 Nano-Omni (23.2% Audio MC) and Qwen3-Omni (24.3% Audio MC) at 54.9% Audio MC; MMAU 77.0% competitive with Qwen3.5 Omni-Plus 81.1% and Gemini 3.5 Flash-Lite 75.2% [§Multimodality, table].
FORTRESS Adversarial 71.6% (Inkling 78.0%, Nemotron 3 Ultra 77.6%, DeepSeek V4 Flash 32.0%) — some regression vs Inkling but still competitive; benign refusal rate 96.9% is best in the compared set [§Safety, table].

Why it’s interesting

The interesting number here isn’t any single benchmark — it’s that Inkling-Small ≥ Inkling on reasoning + agentic + forecasting at ¼ the parameters. The concrete recipe is on-policy distillation from a frontier teacher (Inkling) into a smaller student, followed by targeted agentic-coding RL — which lands the Thinking Machines On-Policy Distillation blog post as an actually-deployed frontier training recipe rather than a research-scale demo on Qwen3-8B. That write-up predicted ~10× compute savings vs RL at matched or better quality; Inkling-Small is a data point that the recipe survives the scale-up. This sits alongside the systematic empirical study in Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe, which catalogs failure modes of OPD/OPSD in current recipes — worth re-reading together. The variable-thinking-effort knob showing uniform Inkling-Small ≥ Inkling at every budget (not just at max effort) is the sharpest signal that the improvement is a real generalization gain, not a benchmark-hacking artifact. Compared to the July release notes, the failure to close the gap on knowledge coverage / factuality is a clean example of the smaller-model tradeoff surfacing exactly where predicted — RL + distillation transfer skills, not stored facts.