Inkling Small: Thinking Machines' Compact Powerhouse That Punches Above Its Weight

Overall Score: 6/10 Inkling Small is a smaller, cheaper model that competes with the normal Inkling in intelligence, but isn’t as competitive with other rivals.

The open-weights model landscape just got a lot more interesting. Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has officially released Inkling Small, a compact yet remarkably capable model that delivers frontier-adjacent performance at a fraction of the size of its larger sibling. Here's everything you need to know about the model, and why it matters.

## What Is Inkling Small? Inkling Small is the smaller, more efficient counterpart to Thinking Machines' flagship Inkling model. It's an efficient open-weights model that achieves comparable performance to Inkling at a quarter of its size: a Mixture-of-Experts transformer with 276B total parameters, 12B active, trained on NVIDIA GB300 NVL72 systems. Like Inkling, it features native reasoning over audio and images, variable thinking effort, a context window of up to 1M tokens, and well-rounded performance across a range of benchmarks.

## Key Specifications at a Glance | Feature | Details | | --- | --- | | Total Parameters | 276B (MoE) | | Active Parameters | 12B per token | | Context Window | Up to 1M tokens | | Modalities | Text, image, audio input → text output | | License | Apache 2.0 | | Reasoning | Yes, with variable/controllable effort | | Weights | Available on Hugging Face |

## Performance: Small Model, Big Numbers Don't let the Small in the name fool you. Inkling Small scores 40 on the Artificial Analysis Intelligence Index, placing it #13 out of 99 models tracked, and well above the median score of 25 for comparable models.

Even more impressively, on several benchmarks, the smaller model actually matches or beats its much larger sibling. Inkling-Small matches or exceeds Inkling on reasoning and agentic tasks. On Humanity's Last Exam it scores 31.6%, ahead of Inkling's 29.7%, and the advantage holds at every thinking budget: Inkling-Small's test-time compute curves sit above Inkling's throughout.

Other notable benchmark results include: - On SWEBench-Verified it exceeds 80% - IFBench: 83.4% (vs. Inkling's 79.8%) - HLE with tools: 46.6% vs. Inkling's 46.0%; GPQA Diamond: 88.3% vs. Inkling's 87.2% These gains come thanks to improvements to the smaller model's pre-training data and recipe.

## Architecture: Mixture-of-Experts Done Right Inkling Small inherits the architectural DNA of its larger sibling. Inkling's MoE design largely follows DeepSeek-V3, with 256 routed experts, 2 shared experts, and 6 routed experts active per token, using a sigmoid router with auxiliary-loss-free load balancing. Thinking Machines interleaves sliding-window and global attention at a 5:1 ratio and uses relative positional embeddings instead of RoPE, reporting better extrapolation to longer sequences.

For its multimodal capabilities, the team chose an encoder-free architecture, feeding audio as dMel spectrograms and images as 40×40 pixel patches through a four-layer hMLP, then processing both jointly with text tokens.

Pretraining ran on 45 trillion multimodal tokens using a hybrid optimizer: Muon for large matrix weights and Adam for other parameters. Post-training was bootstrapped with a small amount of SFT on synthetic data from open-weights models, including Kimi K2.5, followed by large-scale asynchronous RL scaled to over 30M rollouts across two continuous runs.

## Controllable Thinking: A Killer Feature One of Inkling Small's most compelling features is its variable thinking effort: the ability to dial reasoning up or down depending on the task. Its variable thinking effort lets users easily adapt it to target their use case, balancing cost and performance. Across many reasoning and agentic benchmarks, Inkling-Small at max reasoning effort has a strong performance-token tradeoff when compared to open weights models in its weight class.

This is a critical design choice. Sweeping the effort setting from 0.2 to 0.99 traces performance against tokens generated, and Inkling matches Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the tokens. For anyone running a model millions of times inside longer workflows, that curve matters more than a peak score.

## Pricing: Genuinely Affordable The pricing is where Inkling Small becomes especially attractive: - Input: $0.30 per 1M tokens - Output: $1.20 per 1M tokens - Cache Hit: $0.06 per 1M tokens (80% discount) At 92 output tokens per second, it also runs faster than the average model (median: 59 tok/s). It cost just $192.37 to evaluate the full Intelligence Index: an accessible price point for both experimentation and production deployment.

## Why Inkling Small Matters ### 1. It Democratizes High-End AI With open weights, anyone can download, run, and retrain the model. MoE routing keeps inference cheaper and faster than the full parameter count implies. Inkling trades raw power for a balance of cost and performance, unlike top closed models. ### 2. It's Built for Fine-Tuning Thinking Machines has made no secret of their strategy: they want you to own your model, not rent one. With 12B active parameters and controllable thinking effort, it is a natural fit for workloads where cost and latency matter such as coding, using LLMs to grade, or generating synthetic data for other models. Inkling-Small trades factual recall and hard agentic coding for lower cost and latency, which suits workloads like grading, synthetic data generation, and high-volume coding. You can fine-tune it directly on Thinking Machines' Tinker platform, making customization straightforward for organizations. ### 3. It Strengthens the U.S. Open-Weights Ecosystem Inkling is built to loosen the grip of closed frontier labs. It gives US teams an open-weights option to sit alongside Chinese open models like Qwen. The bigger Inkling model already made a splash, debuting at 41 on the Artificial Analysis Intelligence Index, making it the leading open weights release from a U.S. lab, scoring 3 points higher than the previous leading U.S. open weights model, Nemotron 3 Ultra (38), and also beating Gemma 4 31B (29) and gpt-oss-120b (24). Inkling Small extends that footprint to a more accessible weight class. ### 4. Community Adoption Was Immediate Many users praised Thinking Machines for releasing the efficient open-weight Inkling-Small model, highlighting its strong coding performance, 1M context window, and day-0 vLLM support. Optimized checkpoints and community benchmarks are already appearing on major hosting sites, and quantized GGUF versions are already up on Hugging Face for local inference.

## Trade-offs to Consider It's worth being clear-eyed about what Inkling Small isn't. The company is blunt about where the model lands. In its own words, Inkling is not the strongest overall model available today, open or closed. What it offers instead is a combination: native multimodal reasoning, efficient and controllable thinking, and immediate availability on Tinker for fine-tuning. The Artificial Analysis evaluation also notes that Inkling Small is somewhat verbose, generating 130M tokens during Intelligence Index evaluation compared to a median of 100M. If token costs matter to your workflow, tuning the thinking-effort dial will be important.

## The Bigger Picture Inkling Small represents something more than just another model release. It's a statement of strategy from one of the most-watched new AI labs. Inkling is a statement about strategy more than a leaderboard win. Thinking Machines is betting that organizations willing to fine-tune and own a model will get more value than they would from a metered call to a closed frontier model, and Inkling is the base they're offering for that. In a landscape increasingly dominated by closed, API-gated frontier models, Thinking Machines is planting a flag in the ground: you can own capable AI. Inkling Small, with its combination of multimodal input, controllable reasoning, permissive Apache 2.0 license, and genuinely competitive benchmark scores, is arguably the most compelling embodiment of that vision yet. For developers, startups, and enterprises looking for a base model to fine-tune, deploy on their own infrastructure, and integrate deeply into their products, Inkling Small looks like one of the most exciting releases of 2026.

**Where to get it:** Weights are available on Hugging Face under an Apache 2.0 license. Fine-tuning is supported natively via Thinking Machines' Tinker platform, and day-0 support exists in vLLM and other popular inference engines.