Thinking Machines releases Inkling: 975B-parameter MoE model with Apache 2.0 license, first major US open-weight multimo
Thinking Machines Lab released Inkling, a mixture-of-experts model with 975B total parameters and 41B active parameters, trained on 45 trillion tokens across text, images, audio, and video. The Apache 2.0-licensed model supports up to 1M context and debuts alongside Inkling-Small (276B-A12B), marking what observers call the strongest US-based open-weight release to date.
Thinking Machines releases Inkling: 975B-parameter MoE model with Apache 2.0 license, first major US open-weight multimodal release
Thinking Machines Lab released Inkling, a mixture-of-experts transformer with 975B total parameters and 41B active parameters, marking the company's first full foundation model release. The model is available under Apache 2.0 license with full weights on Hugging Face.
Core specifications
Inkling was pretrained on 45 trillion tokens spanning text, images, audio, and video. The model supports a context window of up to 1M tokens in the open-weight release, with API access offering 256K context windows.
According to Thinking Machines, the model "reasons natively over text, images, and audio, and balances cost with performance through efficient and controllable thinking effort." Pretraining began in winter 2025, with a small team adding coding, reasoning, and agentic capabilities from mid-January 2026.
Technical architecture
Community analysis of the release identified several architectural choices:
- Hybrid sliding-window attention with 5:1 local-to-global layer ratio and window size 512
- Relative positional encoding instead of RoPE
- Short convolution layers around attention/FFN streams
- MoE with 2 shared experts (versus the typical 1)
- DeepSeek-style auxiliary-loss-free load balancing
- muP and Muon/weight decay variants
- 8 MTP heads for speculative decoding
Variant and ecosystem support
Inkling-Small, a lighter variant with 276B total parameters and 12B active parameters, is released alongside the main model. Community reviewers report unexpectedly competitive performance from the smaller version.
Day-one infrastructure support includes vLLM, SGLang, Modal, Baseten, Databricks, and Hugging Face, with quantization tooling already available.
Benchmark performance
Artificial Analysis placed Inkling at position 41 on its Intelligence Index, making it the leading US open-weight release. The ranking positions it ahead of Nemotron 3 Ultra (38), Gemma 4 31B (29), and gpt-oss-120b (24), though behind top Chinese open-weight models like GLM 5.2 and Kimi K 2.6 on some benchmarks.
Specific scores from Artificial Analysis:
- GDPval-AA v2 Elo: 1238 (above Kimi K2.6 at 1190 and DeepSeek v4 Flash max at 1189)
- τ³-Banking: 24% (above Kimi K2.6 at 21%, similar to DeepSeek v4 Flash max at 23%)
Design Arena reports Inkling entered Agentic Web App Arena at #9 overall with Elo 1257, ranking it the highest US-based open-weight model for agentic workloads.
Pricing
Pricing details were not disclosed in the release materials. The model is available for immediate fine-tuning on Thinking Machines' Tinker platform.
What this means
Inkling represents the first competitive US-based open-weight alternative to Chinese models like DeepSeek and GLM in the 40B+ active parameter class. The Apache 2.0 license removes commercial restrictions that limit models like Llama, potentially accelerating enterprise adoption. The architectural choices—particularly relative positional encoding at scale and the dual shared expert MoE design—offer researchers new baselines for model development. However, the model still trails top closed models and leading Chinese open-weight releases on pure benchmark performance, positioning it as a strong foundation model rather than a frontier breakthrough.
Related Articles
Nvidia releases Nemotron 3.5 Super VL: open 120B multimodal MoE with 12B active parameters
Nvidia has released Nemotron 3.5 Super VL, an open multimodal Mixture-of-Experts model with 120 billion total and 12 billion active parameters, built for continuously running enterprise agents. Two compact Bolt models for local PCs and edge devices launch alongside it, though the 4 GB variant ships later this month.
OpenBMB's MiniCPM-V 4.7 35B-A3B MoE claims 13+ tokens/s on iPhone 17 Pro Max in 3 GB
OpenBMB has released MiniCPM-V 4.7 35B-A3B, the first mixture-of-experts model in the MiniCPM-V series, with open weights on Hugging Face. The company claims it runs at 13+ tokens/s using 3 GB of memory on an iPhone 17 Pro Max, and that it beats Qwen3.6 35B-A3B on overall score in less inference time in Instruct mode.
NVIDIA releases Nemotron 3.5 Super VL NVFP4: 120B-parameter, 12B-active vision-language model with 1M context
NVIDIA has published Nemotron 3.5 Super VL in NVFP4 quantization on Hugging Face: a 120B-parameter, 12B-active hybrid Mamba-2/Transformer MoE with a 1M-token context window and text, image and video input. NVIDIA reports the quantized checkpoint stays close to BF16 on most benchmarks. It fits on one DGX Spark, one H200, two H100s or one GB200.
DeepSeek V4.1 Flash 1017 launches with 552B CED backbone, 1M context, $0.15/$0.60 per 1M tokens
DeepSeek has released V4.1 Flash 1017, an October 17, 2026 snapshot built on its new Causal Encoder-Decoder architecture. It uses a 552B-parameter sparse MoE backbone with a 1M-token context window, native image understanding, and pricing of $0.15 input / $0.60 output per 1M tokens on OpenRouter.
Comments
Loading...