Pyyan / Compare / Inkling 2 Small vs DeepSeek V4.1 Flash
Inkling 2 Small vs DeepSeek V4.1 Flash
2 of 5Open-Weight Models · verified 14 Nov 2026
TInkling 2 SmallThinking Machines Labcurrent
DeepSeek V4.1 FlashDeepSeekcurrent | Specification | Inkling 2 Small | DeepSeek V4.1 Flash |
|---|
| Summary | A 594B multimodal model with 17B active, trained natively in 4-bit NVFP4, under Apache 2.0. | Twice the parameters of the Flash it replaces, and cheaper to run. |
|---|
| Context | 1M | 1M |
|---|
| Input ($/Mtok) | Weights only | $0.3 |
|---|
| Output ($/Mtok) | Weights only | $1.2 |
|---|
| Licence | Apache 2.0 | MIT |
|---|
| Max output | Not published | 256K |
|---|
| Parameters | 594B total, 17B active | 552B total, 8B / 16B active |
|---|
| Released | 5 Nov 2026, weights 12 Nov | 10 Sep 2026, refreshed 17 Oct |
|---|
| Model ID | thinkingmachines/Inkling-2-Small | deepseek-flash |
|---|
| Category | Open-Weight Models | Open-Weight Models |
|---|
| Official | Thinking Machines Lab ↗ | DeepSeek ↗ |
|---|
Highlighted rows are where these differ.
Inkling 2 Small
- Weights published on Hugging Face on 12 November 2026, a week after a hosted-only launch on Tinker on 5 November
- Takes text and images, writes text, with a context window of up to 1M tokens
- Pretrained, midtrained and post-trained in NVFP4. The main checkpoint needs at least 500GB of GPU memory, such as 2 B300s or 4 H200s
- An MXFP8 checkpoint is also published and needs at least 800GB. The lab says it scores only 1 to 2 points higher
- Local serving recipes for SGLang, vLLM, TokenSpeed and Unsloth. The benchmark scores on the card are the lab's own
Best for self-hosting a large text and image model on a multi-GPU node, or fine-tuning it on Tinker.
Full spec sheet →
DeepSeek V4.1 Flash
- 552B total, 8B active while reading the prompt and 16B while generating, with a 1M token context
- DeepSeek says it beats V4-Pro; V4-Pro requests route to it from 14 September 2026
- MIT licence, native image input, and a KV cache about a quarter the size of V4-Flash's
- Refreshed on 17 October as V4.1-Flash-1017, still MIT, improving on all eleven benchmarks DeepSeek published
Best for cheap agentic coding with images.
Full spec sheet →