Model Drop

Inkling

July 15, 2026 2 min read

Inkling, which Thinking Machines released on July 15, 2026, is the lab's first open-weight model: a 975B-parameter multimodal mixture-of-experts (41B active) built as a base for fine-tuning.

Model: Inkling (thinkingmachines/Inkling)
Maker: Thinking Machines
Ship date: July 15, 2026
Pricing: Free (open weights)
Context: 1M tokens (64K and 256K on Tinker)
Weights: Open (Apache 2.0)
Where you can use it:
- Hugging Face weights, BF16 and NVFP4
- Tinker playground and fine-tuning
When to reach for it:
- A big open base to fine-tune on Tinker
- Text, image and audio input in one model
- Skip it if you want the top open scorer
How it compares:
- Trails GLM-5.2 and Kimi K2.6 on most tests
- Edges past Nemotron 3 Ultra on most tests
What makes it different:
- Trained from scratch on text, image, audio
- Thinking effort is a dial, not a switch
Early reactions: Coverage praised the openness and the candid "not the strongest" pitch, with early testers calling it "pretty good" and architecturally distinct.
Read more: Inkling: Our Open-Weights Model, model card

The benchmarks

Thinking Machines ran Inkling at effort 0.99 and temperature 1.0 against open peers Nemotron 3 Ultra, Kimi K2.6, GLM-5.2 and DeepSeek V4 Pro; rival scores are mostly self-reported, HLE and GPQA Diamond come from Artificial Analysis, SWE-bench Verified uses a bash-only harness for Inkling, and Terminal-Bench 2.1 uses an internal harness.

Thinking Machines' numbers

Humanity's Last Exam (no tools) (expert questions): 29.7%. +3.1 over Nemotron 3 Ultra (26.6%), −6.2 under Kimi K2.6 (35.9%), −10.4 under GLM-5.2 (40.1%).
Humanity's Last Exam (with tools) (expert questions): 46.0%. +8.6 over Nemotron 3 Ultra (37.4%), −8.0 under Kimi K2.6 (54.0%), −8.7 under GLM-5.2 (54.7%).
GPQA Diamond (PhD-level science): 87.2%. +0.5 over Nemotron 3 Ultra (86.7%), −3.9 under Kimi K2.6 (91.1%), −2.3 under GLM-5.2 (89.5%).
AIME 2026 (competition math): 97.1%. +2.9 over Nemotron 3 Ultra (94.2%), +0.7 over Kimi K2.6 (96.4%), −2.1 under GLM-5.2 (99.2%).
SWE-bench Verified (real GitHub fixes): 77.6%. +6.9 over Nemotron 3 Ultra (70.7%), −2.6 under Kimi K2.6 (80.2%), −3.0 under DeepSeek V4 Pro (80.6%).
SWE-bench Pro (hard real-repo fixes): 54.3%. +7.9 over Nemotron 3 Ultra (46.4%), −4.3 under Kimi K2.6 (58.6%), −7.8 under GLM-5.2 (62.1%).
Terminal-Bench 2.1 (agentic terminal tasks): 63.8%. +7.4 over Nemotron 3 Ultra (56.4%), −7.5 under Kimi K2.6 (71.3%), −18.9 under GLM-5.2 (82.7%).
IFBench (instruction following): 79.8%. −1.6 under Nemotron 3 Ultra (81.4%), +3.8 over Kimi K2.6 (76.0%), +6.5 over GLM-5.2 (73.3%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 25.0. +2.1 over Nemotron 3 Ultra (22.9), −2.0 under Kimi K2.6 (27.0), −5.4 under DeepSeek V4 Pro (30.4), −8.7 under GLM-5.2 (33.7).