Inkling-Small: The 276B Open-Weight MoE That Outperforms Its 975B Teacher on Reasoning and Coding

Benchmarks: Where Inkling-Small Wins — and Where It Doesn't

Inkling-Small outperforms Inkling on three of the most demanding public benchmarks, but trails on factual recall and competition math .

Sources:

Key takeaways:

• Reasoning: Inkling-Small scores 31.6% on Humanity's Last Exam (text-only), ahead of Inkling's 29.7%. The advantage holds across thinking budgets .
• Coding agents: 80.2% on SWE-Bench Verified (bash-only harness, 256K trajectory) — a gain of 2.6 percentage points .
• Science: 89.5% on GPQA Diamond, outperforming Inkling's 87.2% .
• Factuality gap: The smaller model trails significantly on SimpleQA Verified (20.6% vs 43.9%), a known trade-off when distilling for reasoning rather than memorization .
• Third-party index: Inkling-Small scores 40 on the Artificial Analysis Intelligence Index v4.1, within one point of Inkling's 41 .

Architecture: How a 276B MoE with 12B Active Works

Inkling-Small is a 42-layer sparse MoE transformer with 6 out of 256 experts activated per token, plus 2 shared experts . The model uses hybrid attention (a mix of full and sliding-window attention) and supports controllable thinking effort, letting developers trade latency for reasoning depth .

The architecture is the same MoE family as Inkling but with fewer total experts (256 vs. Inkling's roughly 576) and fewer active experts per layer. The result is a model that behaves like a 12B-parameter dense model at inference time but retains the capacity of a 276B model in its weights .

Two-Step Post-Training Process

Thinking Machines used a two-stage recipe that is now being studied as a standard post-training primitive in the industry :

1. On-policy distillation from Inkling — An earlier checkpoint (Inkling-Small preview) was trained using Inkling as the teacher. The student sampled its own trajectories while the teacher provided dense token-level supervision — a method known as on-policy distillation (OPD) . This technique combines the sample efficiency of distillation with the distributional realism of on-policy training.

2. Two additional weeks of agent-focused reinforcement learning — The team continued scaling agentic coding RL, which pushed the model past Inkling on reasoning and coding benchmarks .

This recipe is notable because the student surpassed its teacher on several tasks after only two weeks of additional RL — a result that aligns with recent research on weak-to-strong generalization via on-policy distillation .

Hardware Requirements: A ~3x VRAM Reduction

Inkling-Small drastically lowers the hardware bar for open-weight models. Below are the official configurations from the model card :

For context, Inkling's BF16 checkpoint required ~1.9 TB of aggregated VRAM . The official NVFP4 quantized variant of Inkling required ~600 GB .

This 3x VRAM reduction makes Inkling-Small the first model in the Thinking Machines lineup that is realistically accessible for LoRA and full-parameter fine-tuning by mid-sized teams without massive multi-node clusters . The model supports BF16, MXFP8, and NVFP4 numerics .

What This Means for Fine-Tuning

• Full-parameter fine-tuning in BF16: Requires a cluster of 4× B300 or 8× H200 GPUs — a significant investment but within reach of well-funded startups and university labs .
• LoRA fine-tuning in NVFP4: Can be done on 2× H200 GPUs, massively reducing the barrier to customization .
• Inference in NVFP4: Fits on a single B300 GPU in W4A4 mode — the first Thinking Machines model that can be run on a single high-end data-center GPU .

Availability and Cost

Inkling-Small is available in several ways :

• Full weights: Available on Hugging Face under Apache 2.0.
• Tinker Playground: Chat with the model using text, image, and audio.
• Tinker API: Fine-tune the model using LoRA. Output pricing is $1.20 per 1 million tokens .
• Tinker API fine-tuning: LoRA fine-tuning is available through the Tinker platform .

Leadership Context

Thinking Machines Lab was founded by former OpenAI CTO Mira Murati after she left OpenAI in 2024. John Schulman (former OpenAI co-founder of the RLHF team) is also a co-founder. Lilian Weng — who was initially part of the founding team — has since departed Thinking Machines Lab and returned to OpenAI. This leaves Mira Murati and John Schulman as the remaining co-founders of the startup .

The Bottom Line

Inkling-Small is a strong statement from Thinking Machines Lab: with the right post-training recipe, a model one-quarter the size can not only match but exceed its larger teacher on reasoning and agentic coding tasks. For developers and organizations that were priced out of running Inkling (1.9 TB VRAM), Inkling-Small opens the door at 600 GB (BF16) or even 180 GB (NVFP4). The trade-off is a significant hit to factual recall, but for coding, reasoning, and instruction-following workloads, Inkling-Small is the more practical open-weight choice as of July 2026.