Inkling-Small
One-line: Thinking Machines’ second open-weight model — a 12B-active MoE that matches or beats its 41B-active sibling Inkling on SWE-bench Verified (80.2 vs 77.6) while being ~1/4 the size; natively multimodal (text/image/audio in), Apache-2.0. ⚠ large (276B total, NVFP4 ~140GB).
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | not published |
| HumanEval+ | — | not published |
| MBPP | — | not published |
| LiveCodeBench | — | not published |
| Aider polyglot | — | not published |
| SWE-bench Verified | 80.2 | lab (bash-only harness) |
| SWE-bench Multilingual | — | not published |
| SWE-bench Pro (public) | 55.9 | lab |
| Terminal Bench 2.1 | 64.7 | lab (internal coding harness; contaminated web-search solutions zeroed) |
| SciCode | 48.7 | lab |
SWE-bench Verified 80.2 beats Qwen3.5-397B-A17B (76.4), MiniMax M2.7 (79.9), DeepSeek V4 Flash (79.0) and sibling Inkling (77.6) in the card’s own comparison table; within ~2 points of DeepSeek-V4-Flash-0731’s class. Lab-reported (bash-only harness per card footnote); echoed by benchlm and VentureBeat coverage.
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | not published |
| τ-Bench | — | not published (Tau 3 Banking: 15.5 lab — weak vs DeepSeek V4 Flash 22.9) |
| ToolACE | — | not published |
| GAIA | — | not published |
| MCP Atlas (public/all) | 79.6 / 79.2 | lab |
| Toolathlon Verified | 54.4 | lab |
| BrowseComp (with context mgmt) | 77.4 | lab |
MCP Atlas 79.6 is the strongest open-weight number in the card’s table (sibling Inkling 78.8, DeepSeek V4 Flash 69.0). Tool-use is a clear strength; multi-turn reliability (τ-Bench banking) is the weak spot.
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU / MMLU-Pro | — | not published |
| GPQA Diamond | 89.5 | lab |
| MATH | — | not published |
| AIME 2026 | 95.5 | lab |
| HLE (text / with tools) | 31.6 / 47.8 | lab |
| HMMT Feb 2026 | 90.2 | lab |
| AA Index (v4.1) | 40.0 | lab (Artificial Analysis index in card) |
Efficiency
- Active params: 12B (per token — 6 of 256 routed + 2 shared experts; 42-layer hybrid local/global attention)
- Total params: 276B (full disk-resident weight footprint)
- Context length: 1,048,576 tokens (1M)
- VRAM (fp16 rough): ~552 GB full precision — NOT realistic locally. Official NVFP4 release (thinkingmachines/Inkling-Small-NVFP4) ≈ 140 GB; BF16 variant is for multi-GPU serving (SGLang/vLLM recipes provided). ⚠ large — single-node multi-GPU or heavy quant required.
What makes it notable
Inkling-Small is the efficiency story of late July 2026: at 12B active it matches or beats its 41B-active sibling Inkling on SWE-bench Verified (80.2 vs 77.6), Terminal-Bench 2.1 (64.7 vs 63.8) and HLE (31.6 vs 29.7), and tops the open-weight field on MCP Atlas (79.6). It’s Apache-2.0, natively multimodal, 1M context, with first-class SGLang/vLLM/Unsloth recipes — one of the strongest agent-capable open weights per active-param. Caveats: 276B total means NVFP4 ~140GB to run locally, τ-Bench banking is notably weak (15.5), and all headline numbers are lab-reported from Thinking Machines’ own harness.
See also
- welcome
- Sibling: Inkling (thinkingmachines/Inkling, 41B active, created 2026-07-14)
- Source: https://huggingface.co/thinkingmachines/Inkling-Small