Inkling-Small Arrives: A Quarter the Size, Yet Ahead of Its Parent on Reasoning and Coding

Thinking Machines Lab has released Inkling-Small, its second model. It is a Mixture-of-Experts design with 276B total parameters and 12B active, roughly a quarter the size of Inkling, the flagship that arrived in mid-July. Against the assumption that smaller means weaker, it outscores its parent on reasoning and coding benchmarks. The weights ship under Apache 2.0, and the quantized checkpoint runs on a single GPU. Factual knowledge, however, clearly regressed.

A quarter the size, but the same tier of intelligence

Inkling-Small is a 42-layer decoder-only transformer with a sparse MoE (Mixture of Experts) feed-forward backbone. It carries 276B total parameters, of which only 12B are active per token. Each token routes to 6 of 256 experts, plus 2 shared experts that fire on every token. Training ran on NVIDIA GB300 NVL72 systems.

The flagship Inkling holds 975B total and 41B active, so Inkling-Small is about one-third the size on both counts. Even so, the independent evaluator Artificial Analysis placed Inkling-Small at 40 on its Intelligence Index, one point behind Inkling's 41. No open-weights model at that size or smaller scores higher, according to the same analysis.

One detail worth checking before you build on it: the official blog lists a context window of up to 1M tokens, while Artificial Analysis records 256K. The ceiling appears to differ by delivery path.

Thinking Machines Lab is the San Francisco research lab founded by former OpenAI CTO Mira Murati. This second model landed roughly two weeks after Inkling.

Training later turned out to be an advantage

The smaller model passed its parent because of the order in which they were trained. According to the company, Inkling-Small started training after the larger model, which left room to revise the pre-training data mix and the machine learning recipe.

On top of that, an intermediate checkpoint was post-trained in part through on-policy distillation with Inkling as the teacher. From there the team spent two more weeks scaling agentic coding reinforcement learning. The gains came from reworking the training procedure rather than from adding parameters.

Ahead where it counts, behind on knowledge

Line up the numbers and the split is stark.

On reasoning, Humanity's Last Exam (text only) comes in at 31.6 percent against Inkling's 29.7 percent. GPQA Diamond is 89.5 versus 87.2, and ARC-AGI-2 is 40.1 versus 36.5. On coding, SWEBench Verified is 80.2 versus 77.6 and Terminal Bench 2.1 is 64.7 versus 63.8. Toolathlon Verified, which measures tool use, shows a gap of nearly 10 points at 54.4 versus 45.5.

What fell off is factual recall. SimpleQA Verified drops to 20.6 percent, less than half of Inkling's 43.9 percent. The financial task τ³-Banking also trails at 15.5 versus 23.7. In the Artificial Analysis tally, the AA-Omniscience index sits at minus 9.0 against Inkling's 2.1, meaning wrong answers outnumber right ones. The evaluator notes this is not caused by more hallucination but by a low accuracy rate of 31 percent; the hallucination rate is in fact lower than Inkling's, at 57 percent.

In short, the reasoning and tool-use abilities survived the shrink while the stored knowledge did not. If your design supplies facts from outside through search or RAG, the impact is limited. If you lean on the model's own memory, it matters.

Frugal with output tokens

Token consumption is worth noting too. Inkling-Small averaged about 24,000 output tokens per Artificial Analysis Intelligence Index task, slightly under Inkling's roughly 25,000. Peers at the same intelligence tier ran far longer: DeepSeek V4 Flash (max) averaged about 45,000 and GPT-5.4 mini (xhigh) about 78,000.

On AA-Briefcase, which measures long-horizon work, it reached 917 Elo against Inkling's 839. Artificial Analysis tempers that result by pointing out that rubric pass rates are nearly identical at 20 versus 19 percent, suggesting better presentation rather than substantially more correct output. It also finished tasks in 34 turns on average, less than half of Inkling's 81.

Audio and vision hold, self-hosting starts at 180GB

Multimodal performance stays close to Inkling. The encoder-free architecture is unchanged: images are split into 40x40 pixel patches and transformed through a four-layer hMLP, while audio is handled as dMel spectrograms. MMMU Pro is 74.0 percent, and CharXiv RQ for chart reading is 77.4 percent, rising to 81.3 percent when the model uses Python to crop and zoom. On the audio side, MMAU is 77.0 percent and VoiceBench is 90.1 percent.

Self-hosting requirements are published as well. The model card states the BF16 checkpoint needs at least 600GB of aggregated VRAM, which means four NVIDIA B300s or eight H200s. The NVFP4 checkpoint lowers that floor to 180GB, running W4A4 on a single B300 (SM100 or later) or W4A16 on two H200s. Supported runtimes include SGLang, vLLM, TokenSpeed, Unsloth and Hugging Face.

Bringing a 276B-class model within reach of one GPU is a meaningful shift. That said, it still assumes access to a rented B300, so this is not something you run on a laptop.

Safety results and how to get it

Safety post-training follows the same recipe as Inkling, with the same pre-deployment process of internal evaluations and red-teaming by external partners. StrongREJECT, which checks refusal of unambiguously harmful requests, comes in at 98.4 percent. On FORTRESS, covering crime, violence and dual-use risk, the adversarial refusal rate is 71.6 percent and the benign answer rate is 96.9 percent. The lab concludes the model offers no material uplift over the existing open-weights ecosystem, while recommending downstream moderation for consumer-facing deployments.

The weights are distributed on Hugging Face under Apache 2.0. The model can also be fine-tuned on Tinker, the company's tuning platform, and tried through text, image and audio chat on Tinker Playground. Both Inkling and Inkling-Small carry a limited-time discount.

Summary

Inkling-Small pairs 276B total parameters with 12B active and still beats a parent nearly four times its size on reasoning and coding. The weights are Apache 2.0, and the NVFP4 checkpoint runs in 180GB of VRAM. On the other hand, factual scores fell by more than half, so it is a poor fit for workloads that rely on the model's memory alone. Plan on pairing it with search and tools, and measure it on your own tasks before deciding.