Inkling 2 Small Review: How Good Is Thinking Machines' MoE?

How good is Inkling 2 Small?

Inkling 2 Small is a strong, uneven model. Thinking Machines built it as a mixture-of-experts system with 594 billion total parameters and 17 billion active per token. It reads text and images. In hands-on tests it handled front-end code generation, multilingual output, photo reading and hand-drawn circuit analysis well. In published benchmarks it trails DeepSeek V4.1 Flash on coding and agent-style tasks, and leads on vision, knowledge work and finance. It is a poor fit as a plain chat model.

What is Inkling 2 Small?

Inkling 2 Small is the latest model from Thinking Machines. It uses a mixture-of-experts design: only 17 billion of the 594 billion parameters activate for any given token, which keeps per-token compute far below the total size. The lab pitches coding and long-running agent work as the main focus, but the model is general purpose.

Per a community build card, each MoE layer holds 256 routed experts plus two shared experts. The same card says vision and audio towers exist in the architecture. The video review, however, says the lab lists audio as unsupported in this release, so treat audio as unconfirmed for Small.

How did it do in hands-on tests?

A creator ran five tests, three through the lab's console with an API key. Credits were required, and the narrator called the model expensive.

Donor kebab HTML simulation

The prompt asked for a single self-contained HTML file with no libraries and no external assets. The output showed a vertical heating element with flames on one side and a spinning meat cone on a spit on the other, with shading and a background glow. The flames and rotation looked good. The cone shape and meat texture needed work. This is a useful test because it forces the model to produce procedural graphics from code alone.

Multilingual proverb table

Given one English proverb ("don't count your chickens before they hatch"), the model returned a clean, searchable table across many languages. It covered several scripts and writing directions, including low-resource and regional languages. Most rendered correctly. The narrator spot-checked a handful (Hebrew, Hindi, Urdu, Saraiki) and found them convincing, but a native speaker is the real judge. Nobody has formally scored naturalness here.

Hand-drawn circuit simulator

This was the hardest test. The model received a photo of a hand-drawn circuit and had to simulate it, then fix errors in the drawing. It produced a working simulator with an input frequency slider, a live waveform and a magnitude plot. Gain and phase updated as the slider moved. It also generated a separate tab that listed the errors it found, why each mattered and a fix. It added design checks for the parts that were correct, which nobody requested.

One caution. In the demo, the tool reported a 75 Hz cutoff for 15 kilo-ohm resistors with 20 nF and 10 nF capacitors. The narrator left numerical accuracy to domain experts. If you rely on output like this, check the values by hand.

Wine-photo vision

Thinking Machines shows a photo of three glasses of red wine. The model must name the grape and vintage for each. It got all three right in 19 of 20 trials. This is a lab-published demo, not an independent test, but it explains why vision is central to this release.

Autonomous fine-tuning

The model was told to train a small model to forecast electricity demand, wind power, air quality and Wikipedia page views. It wrote its own evaluator, built training data, ran fine-tuning, tried reinforcement learning and debugged along the way. After about four hours and $426, the forecasting score more than doubled. The narrator's takeaway is that you could hand an agent this job without scripting each step. A fair reading is narrower: one curated task worked, and the cost is real.

How does it compare with DeepSeek V4.1 Flash?

The published tables show a split result.

On Terminal Bench the narrator quoted 13.1% against 33.8%. The spoken phrasing was ambiguous about which score belongs to which model, but the conclusion was clear: DeepSeek is well ahead on that agentic coding test. The narrator also called DeepSeek the stronger model overall, while noting that Inkling is competing with it.

If your workload is terminal-driven coding agents, the benchmarks point to DeepSeek. If it involves reading images, documents or financial material, Inkling is worth testing.

What are its known weaknesses?

The lab lists its own limits, and the reviewer agreed with them:

• It is not the most concise communicator.
• It has no audio support in this release.
• It is less polished as a chat model.

The last point matters for product builders. Multi-turn question and answer, where each reply builds on the last, is not where this model shines. It fits better as a task-oriented worker: generate this file, analyze this image, run this job.

Can you run Inkling 2 Small locally?

Possibly, with heavy hardware. The reviewer described access through the lab's console and hoped for an open release. A community build card, though, links the base model on Hugging Face as thinkingmachines/Inkling-Small, says it inherits an Apache-2.0 license, and notes that Thinking Machines ships an NVFP4 checkpoint. Check the official repos for current license and availability before planning around either.

That community card, from PipeNetwork, describes MLX builds for Apple Silicon:

The pruned build fits a 128 GB Mac, where the unpruned one does not. The card reports no measurable quality loss from the 25% prune: text perplexity moved by −8.4% on one held-out set and +0.55% on another, which the author reads as noise. Vision scored 6/6 and speech overlap 0.896 on its small held-out tests. A 50% prune roughly doubled perplexity, so the author did not publish it.

Two practical notes from the card. The inkling_mm_model architecture is not in stock mlx-lm or mlx-vlm, so you need the bundled loader. And the card's own multimodal profiling found about 27% of experts serve mainly image or audio input, which is why text-only pruning can quietly damage other modalities. These are one third party's measurements on small test sets, not independent benchmarks.

Is Inkling 2 Small worth trying?

It is worth a test if you build vision-heavy, document-heavy or long-running agent workflows and can absorb API costs. The demos show real range: procedural graphics, multilingual structure, technical image analysis and a multi-hour autonomous job. Skip it if you need cheap chat, audio input or the best coding-agent scores, since DeepSeek V4.1 Flash leads there by the published numbers.

The sensible approach is to run your own prompts on a small budget before committing. Every result above comes from curated demos or one reviewer's tests, and none is a controlled evaluation.