Inkling 2 Small, in numbers.
Everything Thinking Machines Lab has published about their second-generation model — specifications, prices, benchmark scores and release status — collected in one place and re-checked every six hours.
Inkling 2 Small is a 594-billion-parameter Mixture-of-Experts model from Thinking Machines Lab, released on Nov 5, 2026, with 17B parameters active per token, native image understanding, and a context window advertised at up to one million tokens. It is currently available through Tinker under the Tinker Terms of Use; the company says Apache 2.0 weights are coming around Nov 12, 2026.
- Total parameters
- 594B
- Active per token
- 17B
- Context, advertised
- 1Mtokens
- Context, served
- 64K / 256Ktokens
- Output price
- $1/ 1M tok
- Intelligence Index
- 31rank 4 / 29
What changed, and what is pending
Full changelog Weights are not out yet. Available now through Tinker for inference and fine-tuning under Tinker Terms of Use. Full weights are promised under Apache 2.0 around Nov 12, 2026 — a window they described as the week after the Nov 5, 2026 launch. Follow the tracker for what would have to change.
- Initial snapshot
Selected benchmarks
All 71 rows across 13 models| Benchmark | Inkling 2 Small | GLM-5.3-Flash | DeepSeek-V4.1-Flash-1017 | Muse Spark Open | Kimi K3 |
|---|---|---|---|---|---|
| FrontierCode v2.0 | 25.4% | 14.2% | 28.5% | 7.6% | 17.3% |
| DeepSWE v1.1 | 72.1% | 63.4% | 77% | 60% | 68.5% |
| Terminal Bench Science v0.1 | 5.7% | 4.8% | 12.9% | 1.9% | 8.1% |
| Terminal Bench v4.0 | 13.1% | 32.8% | 33.8% | 8.1% | 12.6% |
| Terminal Bench 2.1 · Inkling 2 Small: Pi harness | 85.1% | 84.3% | 78.4% | 76.4% | 88.3% |
| Terminal Bench v2.1 (Terminus) | 81.3% | 84.3% | – | 76.4% | 85% |
| Vibe Code Bench v1.1 | 81.6% | 30.8% | 87.7% | 76.2% | 85% |
| Sudo L7 | 14.1% | – | – | – | 13.9% |
| SWE-Marathon v1.1 | 38.8% | 38.8% | 40% | – | 48.1% |
| SWEBench Multilingual | 81.2% | 88.9% | 82.9% | 77.5% | 91.4% |
| SWEBench Multimodal v2 | 43.1% | – | – | – | 44.2% |
| SWEBench Pro v2 (Hard) | 86.3% | – | – | 82.4% | 88.2% |
| SWEBench Pro v2 (Full) | 97.7% | – | – | 95% | 97.7% |
| SWEBench Verified | 82.2% | 92% | 84.6% | 80% | 93.4% |
Price
Pricing detail| Variant | Prefill | Cached prefill | Sample | Train |
|---|---|---|---|---|
| 64K context | $0.3 | $0.03 | $1 | $1.5 |
| 256K context | $0.3 | $0.03 | $1 | $3 |
At a glance
- Architecture
- 48 layers, sparse MoE routing each token to 8 of 384 experts; hybrid local/global attention
- Vision
- Native image input via a 292M-parameter vision encoder, 20×20-pixel patches with 2×2 spatial pooling. Output is text only.
- Numerics
- MXFP8, NVFP4
- License now
- Tinker Terms of Use
- License intended
- Apache 2.0 · targeted Nov 12, 2026
- Speed, third-party
- 135 tok/s output, 34.15s to first token, 18.5 minutes per Intelligence Index task (median for comparable models: 7)