Overview
Inkling-Small is Thinking Machines Lab's efficient sibling to Inkling, released on July 30, 2026 as an Apache-2.0 open-weights model. It is a sparse Mixture-of-Experts transformer with 276 billion total parameters and 12 billion active per token — about a quarter of Inkling's size — and Thinking Machines says it "matches or exceeds Inkling on reasoning and agentic tasks" while trailing on knowledge coverage and factuality.
The model keeps Inkling's design choices: a 42-layer decoder-only transformer with 256 routed experts plus 2 shared experts (6 active per token), hybrid local and global attention, native reasoning over text, images, and audio through an encoder-free architecture with a light-weight embedding layer, variable thinking effort, and a context window of up to 1 million tokens. BF16 and NVFP4 (Blackwell-tuned) checkpoints are both published, so Inkling-Small runs on much smaller boxes — the NVFP4 build fits in a single NVIDIA B300 or two H200 GPUs (~180 GB VRAM), versus the ~600 GB the BF16 Inkling checkpoint needs.
Full weights are on Hugging Face under Apache-2.0. Inkling-Small is served on the Tinker inference platform "with a limited-time discount" (Tinker output pricing is $1.20 per 1M tokens versus $4.05 for Inkling) and is exposed for chat over text, image, and audio through the Tinker Playground. Together AI lists it as an inference provider on the Hugging Face page.
| Released | 2026-07-30 |
|---|---|
| License | Apache-2.0 |
| Weights | Open weights |
| Parameters | 276B total · 12B active |
| Context | 1M |
| Architecture | 42-layer decoder-only transformer, sparse Mixture-of-Experts (256 routed + 2 shared experts, 6 active per token), hybrid local/global attention |
| Modalities | Text, Vision, Audio |
| Status | Generally available |
Benchmarks
Inkling-Small benchmark comparison as published on the Thinking Machines model card, vs open-weight peers and select frontier models.
| Benchmark | Inkling-Small | Inkling | Qwen3.5 397B-A17B | MiMo V2.5 | MiniMax M2.7 | DeepSeek V4 Flash | Nemotron 3 Ultra | Claude 4.5 Haiku | Gemini 3.5 Flash-Lite | GPT 5.6 Luna |
|---|---|---|---|---|---|---|---|---|---|---|
| AA Index v4.1 | 40% | 41% | 34% | 37% | 38% | 40% | 38% | 30% | 36% | 49% |
| SWEBench Verified | 80.2% | 77.6% | 76.4% | 71% | 79.9% | 79% | 70.7% | 73.3% | 75% | 93% |
| Terminal Bench 2.1 | 64.7% | 63.8% | 51.3% | 63.7% | 55.4% | 61.8% | 56.4% | 44.2% | 54% | 82.5% |
| SciCode | 48.7% | 46.1% | 42% | 43.1% | 47% | 44.9% | 39.9% | 43.3% | 40.9% | 50% |
| GPQA Diamond | 89.5% | 87.2% | 89.3% | 84.9% | 87.4% | 89.4% | 86.7% | 67.2% | 83.8% | 89.5% |
| AIME 2026 | 95.5% | 97.1% | 93.3% | 93.6% | 87.7% | 95.8% | 94.2% | 85.1% | 82.2% | 97.6% |
| HMMT Feb 2026 | 90.2% | 86.3% | 87.9% | 82.6% | 71.2% | 93.9% | 78.8% | 66.7% | 63.6% | 98.5% |
| HLE (text only) | 31.6% | 29.7% | 27.3% | 25.2% | 28.1% | 32.1% | 26.6% | 9.7% | 17.5% | 35.6% |
| HLE (with tools) | 47.8% | 46% | 48.3% | 40% | 40.3% | 45.1% | 37.4% | 17.8% | 42.5% | 48.9% |
| IFBench | 82.2% | 79.8% | 78.8% | 67.1% | 75.7% | 79.2% | 81.4% | 54.3% | 78.6% | 67.3% |
| Global-MMLU-Lite | 86.7% | 88.7% | 90% | 83.5% | 83.9% | 88.4% | 85.6% | 83.4% | 89.4% | 88.7% |
This model's scores
- AIME 202695.5%
- GPQA Diamond89.5%
- HMMT Feb 202690.2%
- SWE-Bench Verified80.2%
- Terminal Bench 2.164.7%
- HLE (text only)31.6%
- HLE (with tools)47.8%
- MMMU Pro (Standard 10)74%
- VoiceBench90.1%
- IFBench82.2%
- Global-MMLU-Lite86.7%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Output | $1.20 / 1M tokens |
|---|
Output-token pricing on Tinker with a limited-time launch discount; Inkling output is $4.05 per 1M tokens on the same platform.
Strengths
- Open Apache-2.0 weights, downloadable and commercially usable, at roughly a quarter of Inkling's size
- Matches or exceeds Inkling on reasoning and agentic-coding benchmarks despite the smaller footprint
- Native multimodal input (text, image, audio) with a 1M-token context and variable thinking effort
- NVFP4 checkpoint fits on a single Blackwell B300 or two H200 GPUs — much cheaper to self-host than Inkling
- $1.20 per 1M output tokens on Tinker with a limited-time discount, vs $4.05 for Inkling
Best for
- Reach for it as an open-weight base to fine-tune for a domain, product, or private data on Tinker without renting an Inkling-scale cluster.
- Reach for it when you want native multimodal input (text + image + audio) with a 1M-token context at a lower cost tier.
- Reach for it when a permissive Apache-2.0 license, smaller GPU budget, and self-hostable weights are hard requirements.
How to access
| Provider | Model ID |
|---|---|
| Tinker (Thinking Machines) ↗ | — |
| Together AI ↗ | — |
Inkling — every version
The full lineage of the Inkling line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Inkling-Smallcurrent | 2026-07-30 | 1M | Apache-2.0 |
| Inkling | 2026-07-15 | 1M | Apache-2.0 |
FAQ
What is Inkling-Small?
Inkling-Small is Thinking Machines Lab's efficient open-weights sibling to Inkling, released on July 30, 2026 under Apache-2.0. It is a 276B total / 12B active sparse Mixture-of-Experts transformer with a 1M-token context and native text, image, and audio input — about a quarter of Inkling's size, with comparable performance.
How does Inkling-Small compare to Inkling?
Thinking Machines says Inkling-Small "matches or exceeds Inkling on reasoning and agentic tasks." On the published model card it beats Inkling on SWE-Bench Verified (80.2 vs 77.6), Terminal Bench 2.1 (64.7 vs 63.8), and HLE text-only (31.6 vs 29.7), while Inkling keeps an edge on knowledge coverage and factuality (SimpleQA Verified 43.9 vs 20.6). Inkling-Small is a quarter of Inkling's active-parameter footprint (12B vs 41B) and 3x cheaper per output token on Tinker.
What license does Inkling-Small use?
Inkling-Small is published under Apache-2.0. The BF16 and NVFP4 (NVIDIA Blackwell-tuned) checkpoints are both on Hugging Face at thinkingmachines/Inkling-Small and thinkingmachines/Inkling-Small-NVFP4 and can be downloaded, fine-tuned, and used commercially without extra restrictions.
What hardware do I need to run Inkling-Small?
The BF16 checkpoint needs roughly 600 GB of aggregated GPU VRAM (about 4x NVIDIA B300 or 8x H200 GPUs per Thinking Machines' model card). The NVFP4 build is much smaller — around 180 GB — and fits on a single B300 (W4A4) or 2x H200 GPUs (W4A16).
How much does Inkling-Small cost on Tinker?
Thinking Machines lists Inkling-Small output pricing at $1.20 per 1M tokens on Tinker with a limited-time launch discount, compared to $4.05 per 1M tokens for Inkling on the same platform.