Inkling 2 Small
Thinking Machines·
Overview
Second-generation open-weight model from Thinking Machines Lab: a 594B total / 17B active sparse MoE (48 layers, 8 of 384 experts per token) with a 292M-parameter vision encoder. Text and image input, text output; the native audio input of the first Inkling is gone. API launch on Tinker 2026-11-05 with closed weights; weights published on Hugging Face 2026-11-12 after external safety testing. Two vendor checkpoints, NVFP4 (~394 GB, recommended) and MXFP8 (~623 GB); the vendor states at least 500 GB of aggregate VRAM for NVFP4 (2x B300 or 4x H200), so this is a datacenter model despite the name. The model card metadata declares Apache 2.0, but the repository ships no LICENSE file and the card binds use to Thinking Machines' Model Acceptable Use Policy. Context up to 1M tokens. All vendor evals run at effort 0.99 (xhigh), temperature 1.0.
Capabilities and innovations
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 93% | vendor GPQA Diamond 93.0, effort 0.99 (xhigh), vendor model card (read 2026-11-13). The row is on the vendor's model card page but not in the Hugging Face README table. |
| Humanity’s Last Exam (no tools) | 44.1% | vendor 'HLE (text only)' 44.1, no tools, effort 0.99 (xhigh), recorded for comparability with the no-tools HLE used here. The same table lists HLE-Diamond 28.7, a different subset that is not written into this field. |
| Terminal-Bench 2.x | 81.3% | vendor Terminal Bench v2.1 81.3 with the terminus v2 harness, effort 0.99 (xhigh), vendor model card (read 2026-11-13). The row is on the vendor's model card page but not in the Hugging Face README table. No Vals AI 2.x run exists. |
| Terminal-Bench 4.0 | 12.63% | independent Terminal-Bench 4.0 12.63% (±3.07), measured by Vals AI at effort 0.99, temperature 1 (accessed 2026-11-13). Vals AI is the reference harness for this field. 4.0 is its own generation and not comparable with the 2.x or 3.0 fields: the same model scores 81.3 on 2.1. |
| DeepSWE v1.1 | 72.1% | vendor DeepSWE v1.1 72.1, explicitly labelled v1.1, mini-swe-agent v2.4.6 (the harness DeepSWE fixes), effort 0.99 (xhigh), model card table. No earlier Thinking Machines model appears in the table, so there is no vendor baseline row to check. Competitor rows in the same table are not entered anywhere from this source. Datacurve's own board was not checked. |
| MMMU-Pro | 81.6% | vendor MMMU Pro (Standard 10) 81.6, model card table. |
| ARC-AGI-1 | 96.5% | independent ARC-AGI-1 96.5%, semi-private set, ARC Prize Verified, config XHigh, listed 2026-11-05. Cost per task not reported. |
| ARC-AGI-2 | 81.7% | independent ARC-AGI-2 81.7%, semi-private set, ARC Prize Verified, config XHigh, listed 2026-11-05. Cost per task not reported. |
- SWE-bench Verified: not reported by the vendor
- SWE-bench Pro: not reported by the vendor — The model card reports 'SWEBench Pro v2 (Hard)' 86.3 with mini-swe-agent. That is labelled v2 and is a different benchmark version from the SWE-bench Pro tracked here, so it is not entered.
- MMLU: not reported by the vendor
- MMLU-Pro: not reported by the vendor — Only Global-MMLU-Lite (91.0) is reported, a different benchmark.
- Terminal-Bench 3.0: not reported by the vendor
Architecture and hardware
- Parameters
- 594B total, 17B active per token (MoE)
- Estimated VRAM at Q4
- ~342 GB, Frontier class
- Quantization formats
- NVFP4, MXFP8, GGUF (community), MLX (community)
- Recommended runtime
- SGLang
- License
- Apache 2.0 (model card metadata) + Model Acceptable Use Policy
Links
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.