AI Potluck
Model components / Base / pretrained models

Inkling

Thinking Machines Lab

Thinking Machines Lab's first open-weight foundation model, released 2026-07-15. A natively multimodal MoE (975B total / 41B active) that accepts text, image, and audio inputs with a 1M-token context window, trained on 45T tokens. Apache-2.0 weights on Hugging Face; positioned as a customizable base (via Tinker LoRA) rather than the strongest overall open or closed model. Family also includes Inkling-Small (276B / 12B active), folded into this entry.

Inkling, 975B total / 41B active MoE (6-of-256 + 2 shared experts), hybrid local/global attention, native text/image/audio. Released 15 Jul 2026 under Apache-2.0. Inkling-Small (276B/12B) is a sibling preview in the same family — not a separate map entry. Training data not redistributed.

Openness

3 high confidence
3.0
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)

Apache-2.0 weights are OSI-licensed, but training data and full training code are withheld — open_weights / 3, not open_source / 5. Press often calls this 'open source' because of the license tag; our spectrum requires open data + code for class open_source.

Adoption

3 medium confidence
3.0

HF card shows ~13K downloads in the first days and 1K+ likes; day-0 serving via Tinker plus partners (Together, Modal, Fireworks, etc.). Early but broad distribution for a ~1T open-weight model; level 3 until cumulative usage settles.

Capability

4 high confidence
4.0

Strong 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).

Unchanged since 2026-07-30 (last edited, not re-checked)