
In this article (4)
Inkling 2 Small: 17A-594B in 68 Hours, Efficiency Becomes the Feature
Key Takeaways
- Treat training time as a product signal: faster pretraining can change iteration speed, release cadence, and customization options.
- Compare MoE models by active parameters and routing behavior, not total parameter count alone.
- Wait for open weights and independent evals before making the Muse Spark comparison your procurement strategy.
Why it matters
- ProductProduct teams can evaluate whether fast training and variable effort translate into better agent and coding workflows.
- InvestorsTraining efficiency is becoming a defensible signal because it can affect iteration speed, release timing, and model economics.
Latent Space’s 12T token framing points to a new builder lesson: training speed is now part of the product, not just the lab notebook.
The loudest number in AI this week was not a benchmark score. It was a stopwatch. Latent Space put the spotlight on Thinking Machines Lab’s Inkling 2 Small with the kind of headline that makes GPU schedulers sweat: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark. Strip away the leaderboard confetti, and the useful lesson is sharper: training efficiency is becoming a product feature, like latency, price, or whether your agent quietly eats the Jira backlog at 2 a.m.
The run got loud because it got fast
Latent Space described Inkling 2 Small as being "pretrained from scratch in less than 3 days on <12T tokens native NVFP4," which is the sort of sentence that sounds like a dare written on a cluster invoice. The report framed the run as competitive with, and often ahead of, comparable Chinese models, while emphasizing customizability around Thinking Machines Lab’s broader Tinker direction. That matters because a fast training run changes the practical rhythm of model development: fewer archaeological digs through stale checkpoints, more chances to iterate before the benchmark discourse turns into soup.
Thinking Machines Lab’s own announcement gives the model its headline shape: a Mixture-of-Experts system with 594B total parameters, 17B active, image input support, and variable effort. The company says Inkling 2 Small reaches 31 on the Artificial Analysis Intelligence Index, with gains across agentic coding, vision, knowledge work, and reasoning. Translation for builders: do not stare only at the total parameter count like it is a sacred cow wearing a hoodie. Active parameters, routing, training recipe, and inference controls are doing a lot of the work here.
Inkling 2 Small model shape
- Text and images
- Multimodal transformer
- 48 layer decoder
- Sparse MoE
- 8 of 384 experts
- Text outputs
The architecture is the efficiency argument
The Inkling 2 Small model card says the system is a 48 layer decoder-only transformer with a sparse MoE feed-forward backbone, where each token is sent to 8 of 384 experts. It also says the model accepts text and image inputs and produces text outputs, which puts it in the very practical category of models meant for coding assistants, retrieval systems, chatbots, visual tasks, and agentic workflows. Sparse MoE is not magic, despite sounding like a committee meeting where only eight people actually do the work. It is a compute allocation strategy: keep a huge pool of capacity, but activate only a slice for any token.
That is why 594B total and 17B active are not contradictory, according to the Thinking Machines Lab model card. They are the point. Total parameters describe the available expertise buffet; active parameters describe what your token actually eats. For builders comparing models, the useful question is no longer just, how big is it, but how much useful capacity do I get per unit of training and inference pain?
RL is where the boring miracle hides
Thinking Machines Lab attributes Inkling 2 Small’s gains to improvements throughout training, including "more efficient pre-training and a large, unified RL run." That sentence is doing heavy lifting, and not just because RL is where product demos go to discover gravity. The Decoder reported that Thinking Machines trained the model with reinforcement learning on hundreds of thousands of long-horizon coding tasks, which lines up with the company’s emphasis on agentic coding gains.
This is the part builders should steal conceptually, even if they do not have a heroic compute budget casually lounging in the garage. Fast pretraining gives you a better base sooner, but targeted post-training can decide whether the model can survive multi-step tool use, code repair, visual grounding, or workflow adaptation. If pretraining is teaching a model the language of the internet, RL is making it practice not face-planting during the job interview. Increasingly, competitive models are not winning because of one giant knob. They are winning because every stage of the pipeline is less wasteful.
Open weights are still on the clock
Thinking Machines Lab says Inkling 2 Small will begin on Tinker with closed weights, and that it intends to release full weights under an Apache 2.0 license next week after external safety testing and red-teaming. The Decoder similarly reported that the model is initially proprietary through Tinker, with Apache licensed weights planned to follow. That makes the release simultaneously useful and not fully inspectable yet, which is AI’s favorite suspense genre.
For teams, the immediate move is pragmatic: test it where Tinker access permits, especially on bounded coding, vision, or agent workflows where variable effort and multimodal input actually matter. Do not overfit your roadmap to one reported benchmark or one spicy Latent Space headline, including the Muse Spark comparison, until weights, evals, and deployment behavior are visible enough to poke with a stick. But do pay attention to the pattern. The labs that can train faster can iterate faster, recover faster, and customize faster, which turns efficiency from a research metric into a competitive surface area.
The next thing to watch is not just whether the Apache 2.0 weights arrive on schedule. Watch how quickly independent builders reproduce, fine-tune, stress-test, and cost out the model in real systems. Benchmarks tell you who can sprint in clean shoes. Training efficiency tells you who might survive the mud.
Sources3 sources
The reporting, announcements and research the AI editor worked from. Links open the original publisher.