Inkling Architecture and Benchmark Notes | Sebastian Raschka, PhD

Skip to main content

🌙

Search this website

Submit search

Interesting surprise open-weight LLM drop from Thinking Machines Lab yesterday. Their nearly 1T-parameter Inkling model looks pretty solid on the reported benchmarks.

Compared with GLM-5.2, Inkling performs better on evaluations such as IFBench (79.8% versus 73.3%) and SimpleQA Verified (43.9% versus 38.1%). It performs worse on several reasoning and coding-agent benchmarks, including HLE without tools (29.7% versus 40.1%), SWE-Bench Pro Public (54.3% versus 62.1%), and Terminal-Bench 2.1 (63.8% versus 82.7%).

These are release-time numbers, and some rows combine externally reported values with internal harness results. I would therefore not over-interpret small differences. Overall, Inkling looks like a good all-rounder that is intended for further fine-tuning and specialization. That direction makes sense given that Thinking Machines Lab also develops Tinker, its model customization and fine-tuning platform.

Architecture-wise, Inkling is a 975B-parameter sparse Mixture-of-Experts model with 41B active parameters and a context window of up to 1M tokens.

Figure 1. Inkling architecture and release-time benchmark comparisons. The benchmark panels use results from Thinking Machines Lab's July 15, 2026 announcement. The higher-resolution architecture figure is also available in the LLM Architecture Gallery.

Some side-by-side observations:

Inkling has about 231B more total parameters than GLM-5.2 (975B versus 744B), although their active footprints are almost identical at 41B and 40B parameters.

Inkling is less sparse than Kimi K2.5. It activates 4.2% of its parameters per token, compared with 3.2% for Kimi K2.5 (41B of 975B versus 32B of 1T).

Inkling uses a regular Transformer decoder rather than the hybrid Mamba-Transformer approach used by Nemotron 3 Ultra.

I am curious about token throughput. The larger active footprint than Kimi K2.5 and the use of conventional GQA rather than MLA or a recurrent hybrid stack suggest that raw decoding speed may not be Inkling’s main advantage. However, throughput also depends heavily on quantization, expert parallelism, attention kernels, batching, and hardware. I have not seen directly comparable provider measurements yet.

The overall design follows the recent large-MoE trend, but the architecture has a few interesting surprises:

Small convolution layers in several places. Each decoder layer applies short kernel-4 convolutions after the key and value projections and on the attention and MLP branch outputs. My intuition is that these provide cheap local token mixing and an explicit short-range inductive bias alongside attention.

An additional RMSNorm directly after the token embedding layer. This is separate from the pre-attention RMSNorm inside every transformer block. At first glance, it looks almost redundant, but it is explicitly enabled in the configuration and present in the Transformers implementation. Whether it materially helps would require an ablation.

A learned, input-dependent relative-position bias instead of RoPE. Thinking Machines says that the relative-position approach “performs better and extrapolates better to longer sequences” than RoPE.

Regarding the last point, my intuition is that the sliding-window-heavy architecture helps. Of the 66 decoder layers, 55 use local attention with a small 512-token window. A learned relative-position bias may provide enough positional information within these windows.

In the 11 global layers, the released implementation applies the learned bias only over the preceding 1,024 tokens. Attention beyond that range is effectively content-based with respect to this positional bias. This is somewhat similar to the intuition behind NoPE, which other architectures use in selected global-attention layers.

Overall, Inkling is an interesting variation on the DeepSeek-V3-style MoE recipe and a solid release. Some may find it underwhelming because it does not lead every benchmark. I find the broad and mixed benchmark profile refreshingly honest. It may also indicate that the model is less benchmark-specialized than several recent releases.

And, as always, it is good to see another strong open-weight base model available for fine-tuning and independent study.

Read Next

200,000 Subscribers

Short note celebrating Ahead of AI reaching 200,000 subscribers.

GPT 5.6 Has 72 Possible Configurations. What's A Good Default?

Short note on how GPT 5.6 model and effort choices map onto training-time and inference-time scaling, producing 72 configurations.

Build a Reasoning Model From Scratch Is Out

Short note announcing the release of Build a Reasoning Model From Scratch and linking the publisher and Amazon pages.

RSS

Subscribe via Email

tweet or email. -->

© 2013-2026 Sebastian Raschka