$ cat wiki/models/inkling.md
Inkling
Compared with
Spec
| Attribute | Value |
|---|---|
| Developer | Thinking Machines Lab |
| Released | 2026-07-15 |
| Announced | 2026-07-15 |
| Context window | 1M |
| Pricing | unknown |
| License | Apache 2.0 (open-weight) |
| Availability | Hugging Face (BF16 and NVFP4), Tinker (fine-tuning), SGLang, vLLM, llama.cpp, Unsloth |
Pricing is unknown because Thinking Machines publishes no list price that was | |
| readable from this environment. The Artificial Analysis row carries a | |
Cost per Task USD of -- for Inkling — the publisher's missing value, and their | |
| measured cost of one task rather than a per-token price in any case | |
| (source). |
| Attribute | Value |
|---|---|
| Catalogue id | unknown |
Release Date
2026-07-15 — Thinking Machines Lab's first production model and first open-weights release (source).
Architecture: 975B total parameters, 41B active per token, a multimodal Mixture-of-Experts taking text, images and audio as input and producing text. Pretraining covered 45 trillion tokens of text, images, audio and video. The MoE design largely follows DeepSeek-V3 — 256 routed experts, 2 shared, 6 active per token (source).
The model exposes a controllable "thinking effort" setting (source), which places it in the same product shape as the effort tiers this wiki records for Claude Opus 5 and Kimi K3 — see Test-Time Compute (Inference-Time Compute Scaling).
Benchmarks
No vendor benchmark table was readable from this environment. What this repo holds is third-party: its own snapshot of the Artificial Analysis leaderboard, read 2026-08-02. Columns are the publisher's own (source):
| Model | Reasoning (derived) | Context Window | Creator | Artificial Analysis Intelligence Index | Cost per Task USD | Median Tokens/s | Latency First Chunk (s) | Total Response (s) |
|---|---|---|---|---|---|---|---|---|
| Inkling | yes | 1M | Thinking Machines | 41 | -- | 86 | 1.87 | 31.04 |
| Inkling Small | yes | 1M | Thinking Machines | 40 | $0.07 | 95 | 1.67 | 27.96 |
| For scale on the same table and the same read: Claude Opus 5 (max) is 61, GPT-5.6 | ||||||||
| Sol (max) 59, Kimi K3 (max) 57, DeepSeek V4 Flash 0731 (max) 50 | ||||||||
| (source). |
Two things follow, and only these two. Inkling is not at the capability frontier on this index. It is fast — 86 median tokens/s and 1.87s to first chunk, against 54 and 92.02s for Claude Opus 5 (max) on the same read (source).
Reasoning (derived) is this repo's column, not the publisher's — it comes from an
icon on their page, per Eval Harness Configuration.
Use Cases
The lab's own stated role for the model is a balanced, multimodal, customizable base for downstream fine-tuning, and it explicitly does not claim Inkling is the strongest overall model (source). Same-day fine-tuning on the lab's Tinker platform, and day-zero support in SGLang, vLLM, llama.cpp and Unsloth, are the operational form of that claim (source).
Compared To
- Kimi K3 — 2.8T, the larger open-weight release of the same month; scores 57 on the Artificial Analysis Intelligence Index against Inkling's 41 on the 2026-08-02 read (source)
- Laguna S 2.1 — 118B/8B, the specialist to Inkling's generalist; the two were named together as open models on the Pareto frontier (source)
- DeepSeek V4-Flash — 284B/13B, MIT; 50 on the same index at a far smaller size (source)
- Mistral Large 3 — the other Apache 2.0 frontier-scale weights release this wiki tracks