Thinking Machines' Inkling 2 Small is a 594B MoE model with strong vision and autonomy, but mixed coding benchmarks. Here's what it gets right.
What is Inkling 2 Small? #
Inkling 2 Small is a mixture of experts (MoE) model from Thinking Machines, built with 594 billion total parameters and 17 billion active parameters per forward pass. It handles both text and images, and the lab positions it as its second generation frontier model, with particular emphasis on coding and long running autonomous agent work. It’s also a general purpose model, meaning it’s meant to handle everyday reasoning, multilingual tasks, and knowledge work, not just narrow coding benchmarks.
TL;DR #
- Inkling 2 Small is a 594 billion parameter MoE model with only 17 billion parameters active per inference, keeping compute costs lower than the total parameter count suggests.
- The model supports vision input , and in testing it correctly identified wine varietals and vintages from a single photo in 19 out of 20 trials.
- An autonomy demo had the model independently build a small forecasting model (electricity demand, wind power, air quality, Wikipedia views), writing its own evaluator, generating training data, running fine-tuning and reinforcement learning, and debugging itself, taking about four hours and $426 in compute to more than double the forecasting score.
- Hands-on tests showed strong results in self-contained HTML/SVG generation (a doner kebab animation built from a text prompt with no external assets) and a surprisingly workablecircuit simulation built from a single hand-drawn circuit image.
- Benchmark results are mixed : on terminal-bench style coding tasks, Inkling 2 Small scored 13.1% versus DeepSeek V4.1 Flash’s 33.8%, but Inkling pulled ahead on vision, knowledge work, and finance-related tasks.
- Thinking Machines itself flags the model’s weaknesses : it isn’t the most concise communicator, has no audio support, and is less polished as a straightforward chat assistant.
- The model is currently only available through Thinking Machines’ own console/API , and pricing was described as expensive relative to competitors; a quantized MLX community conversion has also appeared on Hugging Face under an Apache 2.0 license for the base weights.
Rapidlybuild enterprise-grade, production-ready software.
Already in:✓ SOC 2✓ HIPAA✓ Audit log
How big is Inkling 2 Small and what does the architecture mean in practice? #
The “594B total, 17B active” structure is standard mixture of experts design: the full model contains many specialized expert sub-networks, but only a fraction of them fire for any given token. This keeps inference cost and latency closer to a much smaller dense model while still letting the system draw on a large overall parameter pool for capability. It’s the same general strategy used by other large MoE releases, including DeepSeek’s V4.1 Flash, which shows up repeatedly as the comparison point in benchmark tables for this release.
Practically, this means Inkling 2 Small isn’t something most people will run locally on consumer hardware in its original form. A community-produced 4-bit MLX quantization (listed under the name “Inkling-Small-MLX-REAP25-4bit” on Hugging Face) exists for Apple Silicon users, built on a pruned (“REAP”) version of the base weights, but the primary way to access the full model is through Thinking Machines’ hosted console and API.
How good is Inkling 2 Small at vision tasks? #
Vision is where this release draws the most attention outside of coding. In one test, the model was shown a single photograph of three glasses of red wine and asked to name the grape variety and vintage for each glass. It got all three correct in 19 out of 20 attempts, a result that suggests genuinely fine-grained visual reasoning rather than surface-level pattern matching.
That strength held up in further independent testing. Given a hand-drawn circuit diagram, with errors introduced by the original AI-generated sketch, the model was asked to simulate the circuit and fix the mistakes. It produced a working simulation with an adjustable input frequency slider, a live waveform, and a magnitude plot, correctly reporting a cutoff frequency for a filter built from 15 kilo-ohm resistors and 20/10 nanofarad capacitors. It also identified specific errors in the original circuit, explained why each one mattered, proposed fixes, and flagged which parts of the design were already correct, without being asked to do that last part. Benchmark tables back this up: Inkling 2 Small reportedly leads on several vision, knowledge work, and finance-related evaluation categories, even where it trails on general coding benchmarks.
Can Inkling 2 Small really act autonomously on long-running tasks? #
This is the headline capability Thinking Machines wants people to notice. In a demonstrated scenario, the model was given a single high-level task: build a small model to forecast electricity demand, wind power output, air quality, and Wikipedia page view trends. Left to work independently, it wrote its own evaluation criteria, generated its own training data, ran a fine-tuning pass, attempted reinforcement learning, and debugged issues as they came up, all without step-by-step human guidance.
production-readysoftware.
The process took about four hours and cost roughly $426 in compute, and the forecasting accuracy score more than doubled by the end. The implication is that for certain applied ML tasks, a user can hand over a plain-language goal and let the model manage the fine-tuning pipeline itself, including dataset creation, rather than someone manually setting up SFT or RL workflows. That’s a meaningful shift for anyone who finds fine-tuning tooling to be the main barrier to customizing a model, though the dollar cost is a real constraint for casual experimentation.
How does Inkling 2 Small perform on coding benchmarks? #
This is the model’s weakest publicly tested area, and the gap is notable. On terminal-bench style coding evaluations, Inkling 2 Small scored 13.1%, compared to 33.8% for DeepSeek V4.1 Flash, a substantially stronger result for the competing MoE model. On the broader artificial analysis index, DeepSeek V4.1 Flash also leads.
That said, raw coding benchmark scores don’t capture everything. In a hands-on test, the model was asked to generate a self-contained HTML file simulating a rotating doner kebab, with flames, a spit, and a layered meat texture, using no external libraries, image assets, or image generation tools. Working from a single text prompt, it produced a visually coherent animation: a convincing flame shape, reasonable shading on the meat cone, and a believable rotation effect, even if the meat texture itself had room for improvement. That’s a creative-coding result rather than a benchmark score, but it’s a useful signal that the model can handle open-ended generative coding tasks even while lagging on structured coding benchmarks.
How well does it handle multilingual tasks? #
In testing, the model was asked to translate a single English proverb (“don’t count your chickens before they hatch”) into a wide range of languages and return the results as a searchable table. It handled major world languages like Hebrew, Hindi, Indonesian, and Urdu cleanly, with correct script rendering and writing direction. More notably, it also produced plausible translations for lower-resource and regional languages, including Seraiki and Nepali, areas where many models either fail outright or produce noticeably weaker output. The table format itself, clean and searchable, was also called out as a practical, usable design choice rather than a wall of text.
Is Inkling 2 Small worth using right now? #
It depends heavily on the use case. For vision-heavy tasks, STEM reasoning, applied forecasting or fine-tuning workflows, and multilingual generation, the model performs well above what its mixed coding benchmark scores might suggest. For straightforward conversational chat, question-and-answer sessions, or general coding assistance compared to strong alternatives like DeepSeek V4.1 Flash, it’s a weaker choice. Thinking Machines itself acknowledges the model isn’t the most concise communicator, lacks audio support, and isn’t as polished a chat model as some competitors. Combined with availability limited to the company’s own console and pricing on the expensive side, this looks like a model best suited to specific technical workloads rather than a general daily driver.
Frequently Asked Questions #
What is Inkling 2 Small’s parameter count?
It has 594 billion total parameters in a mixture of experts architecture, with only 17 billion parameters active during any given inference pass.
Does Inkling 2 Small support images?
Yes. It accepts both text and image input, and vision is one of its strongest demonstrated capabilities, including detailed tasks like identifying wine varietals from photos and interpreting hand-drawn circuit diagrams.
How does Inkling 2 Small compare to DeepSeek on coding?
regulatory compliance and security.
- Name
- Maria Alvarez
- Date of birth
- SSN
- Diagnosis
- Balance due
- $412.80
It currently trails DeepSeek V4.1 Flash on coding benchmarks like terminal-bench (13.1% versus 33.8%) and on the broader artificial analysis index, though it performs better in vision, knowledge work, and finance-related evaluations.
Can I run Inkling 2 Small locally?
The full model is primarily accessed through Thinking Machines’ own console and API. A community-made 4-bit MLX quantization based on a pruned version of the weights is available for Apple Silicon, under an Apache 2.0 license.
What are Inkling 2 Small’s main weaknesses?
According to the lab itself, it isn’t the most concise communicator, has no audio support, and is less polished as a conversational chat model compared to some competitors.