Thinking Machines Lab выпустила Inkling-Small: открытая MoE-модель 276B превзошла учителя

Thinking Machines Lab on August 2, 2026 released the weights of Inkling-Small — a multimodal model based on the Mixture-of-Experts architecture with 276 billion total parameters and 12 billion active. At four times smaller than the original Inkling (975B total, 41B active), the new model surpasses it on most key benchmarks in coding and reasoning.

How Inkling-Small differs from Inkling

Inkling-Small is a 42-layer decoder-transformer with a sparse MoE layer instead of standard FFN. Each token is routed to 6 out of 256 experts plus 2 shared experts that are always active. Attention is built as a hybrid of local and global layers.

The model is natively multimodal and does not require a separate encoder:

• 276B total parameters, 12B active during inference
• Context window — 1M tokens with an adjustable "effort" level of thinking
• Input: text, images (40×40 pixel patches via a four-layer hMLP) and audio (dMel spectrograms, WAV 16 kHz, up to 2 minutes)
• Numerical formats: BF16, MXFP8 and NVFP4
• Apache 2.0 license, weights published on Hugging Face

How to run Inkling-Small

The BF16 checkpoint requires a minimum of 600 GB of total VRAM — that is 4× NVIDIA B300 or 8× NVIDIA H200. The quantized NVFP4 version lowers the threshold to 180 GB: W4A4 mode runs on a single B300 card (SM100+ architecture required), W4A16 — on two H200s.

Supported runtimes are SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face Transformers. The single-GPU path takes the 276B model out of the "large labs only" category: startups can rent a single B300 instance, and companies with existing H200 clusters avoid new purchases. Regulated industries — finance, healthcare, insurance, telecom — gain a private-weights option without dependence on an external API.

On which benchmarks Inkling-Small outperformed Inkling

Inkling-Small surpassed the "teacher" on most key tests. On Humanity's Last Exam (text version) it scores 31.6% versus 29.7% for Inkling. SWE-bench Verified — 80.2% versus 77.6% (bash-only harness). Toolathlon Verified — 54.4% versus 45.5%. GPQA Diamond — 89.5%, AIME 2026 — 95.5%, ARC-AGI-2 — 40.1% versus 36.5%. Terminal-Bench 2.1 — 64.7%.

Two clear regressions: SimpleQA Verified drops to 20.6% from 43.9% for Inkling, Tau 3 Banking — to 15.5% from 23.7%. Multimodal scores remain close to the teacher: MMMU Pro — 74.0%, VoiceBench — 90.1%, MMAU — 77.0%, according to Artificial Analysis, Scale AI and ARC Prize.

"The model does not create a significant advantage over the existing open-weights ecosystem," summarizes

Thinking Machines Lab in the Inkling-Small model card, recommending adding downstream moderation — such as Llama Guard — in consumer applications.

What this means

Inkling-Small demonstrates that targeted distillation with the Inkling teacher and two weeks of RL training on agentic coding allow a compact model to outperform a larger predecessor. Open weights under Apache 2.0 make the 276B agent available for self-deployment — and expand the pool of companies that can build products without dependence on closed APIs.

Frequently Asked Questions

How much VRAM is needed for Inkling-Small?

A minimum of 180 GB VRAM in quantized NVFP4 mode on a single B300 card (SM100+). Full BF16 precision requires 600 GB total — for example, 4× NVIDIA B300 or 8× NVIDIA H200.

How is Inkling-Small better than the original Inkling?

On six key benchmarks: SWE-bench Verified (80.2% vs 77.6%), GPQA Diamond (89.5%), AIME 2026 (95.5%), ARC-AGI-2 (40.1% vs 36.5%), Toolathlon Verified (54.4% vs 45.5%) and HLE — 31.6% versus 29.7%. It falls short on SimpleQA Verified (20.6% vs 43.9%) and Tau 3 Banking (15.5% vs 23.7%).