Inkling
Thinking Machines LabThinking Machines Lab's first open-weight foundation model, released 15 July 2026. It is a natively multimodal mixture of experts with 975 billion total and 41 billion active parameters, six of 256 experts plus two shared, hybrid local and global attention, a one-million-token context and 45 trillion training tokens. A 276B Small sibling with 12B active is part of the same family.
The Small sibling is folded into this record. Training data is not redistributed. The NVFP4 quantization pulls roughly three times the BF16 release, so adoption sums both rather than reading the headline repository. Verified 2026-08-13 via the Inkling model card and Thinking Machines' launch post.
Openness
3 high confidence- weights
- open(Apache-2.0 on HF, BF16 + NVFP4)
- data
- closed(public+third-party+synthetic, not redistributed)
- code
- partial(day-0 transformers/SGLang/vLLM/llama.cpp
- license
- Apache-2.0(OSI for weights only)
The Apache-2.0 weights are OSI-licensed, but the training data and the full training pipeline are withheld, so this is open weights (3) rather than open source (5). Press coverage often calls Inkling open source because of the license tag; on this map open source also requires the data and the training code.
- https://huggingface.co/thinkingmachines/Inkling recorded 2026-08-13
Apache-2.0 license tag; 975B/41B MoE; weights downloadable; model card states data from public, third-party, and synthetic sources (not released)
- https://thinkingmachines.ai/news/introducing-inkling/ recorded 2026-08-13
primary announcement - weights on HF, Tinker availability, Inkling-Small sibling, and the Together / Modal / Fireworks inference partners
Adoption
3 medium confidence368,909 downloads in the trailing 30 days across the two declared artifacts (thinkingmachines/Inkling 96,344; thinkingmachines/Inkling-NVFP4 272,565), which bands at level 3 (100K-1M) on the model adoption scale. The NVFP4 quantization outpulls the BF16 release almost three to one, which is why the family is summed rather than banded on the headline repository alone.
- https://huggingface.co/api/models/thinkingmachines/Inkling recorded 2026-08-13
96,344 downloads in the trailing 30 days for thinkingmachines/Inkling
- https://huggingface.co/api/models/thinkingmachines/Inkling-NVFP4 recorded 2026-08-13
272,565 downloads in the trailing 30 days for thinkingmachines/Inkling-NVFP4
Capability
4 high confidenceStrong open-weight multimodal base — beats Nemotron 3 Ultra on SWE Verified (77.6 vs 70.7) but trails Kimi K2.6 (80.2) and DeepSeek-V4-Pro (80.6). Vendor card is explicit that Inkling is not the strongest overall model; scored 4 (high open-weight tier, not frontier-defining 5).
- https://huggingface.co/thinkingmachines/Inkling recorded 2026-08-13
vendor eval table vs Nemotron 3 Ultra / Kimi K2.6 / DeepSeek V4 Pro / closed frontier peers, carrying SWE-bench Verified 77.6, SWE-bench Pro 54.3, GPQA Diamond 87.2 and AIME 2026 97.1
- https://thinkingmachines.ai/model-card/inkling/ recorded 2026-08-13
model card with architecture and evaluation details
Verified 2026-08-13