AI Desk · AI FingerprintingWithdrawn *
Why high: an exact fingerprint shared with 2 models that declare the same family, without a second signal of a different kind yet.
DeepSeek: R1 Distill Llama 70B is DeepSeek's model, and its measured tokenizer sits in DeepSeek's own signature group alongside its sibling models. Measured models from StepFun carry the same vocabulary, which reads as reuse of DeepSeek's tokenizer line and questions nobody's authorship; they are listed under "Shares this fingerprint" below.
| The weights' own tokenizer | Llama 3, verified offline against the model's repository: 50 of 50 strings, vocabulary 128,256 |
| The endpoint's token accounting | DeepSeek's vocabulary: the billed counts match DeepSeek-V3's tokenizer on every string once the one-token template boundary is folded, and the model's own tokenizer on only 18 of 50 |
| Where the two meet | The bill follows the serving layer rather than the weights, so the catalogue's Llama3 label stays right about the model itself |
| Not established by this | That the weights run a DeepSeek vocabulary; the repository tokenizer says they do not |
The mismatch was traced to the serving layer, not the label. The model's own repository tokenizer, run over the same 50 strings offline, is Llama 3's exactly (50 of 50, vocabulary 128,256), so the catalogue's Llama3 tag is right about the weights. The endpoint's reported token counts tell a different story: they match DeepSeek-V3's repository tokenizer on every string once the template boundary is folded, and the model's own tokenizer on only 18 of 50. The token accounting behind this endpoint uses a DeepSeek vocabulary the model does not read with, which sets what users are billed. Whether that is a billing-side tokenizer configuration or something else about the serving path cannot be established from token counts alone.
DeepSeek-R1-Distill-Llama-70B repository (config.json: vocab_size 128256, LlamaForCausalLM)DeepSeek-V3 repository (config.json: vocab_size 129280)
| We checked | What we found | Where it came from |
|---|---|---|
| How it counts tokens | Counts every one of the 50 test strings exactly like 2 other models | tk_869dc1cf |
| How its API is set up | A combination of settings no other listing in the catalogue uses | api_3560b58d |
| How much it can read and write | Reads up to 8,192 tokens at onceOpenRouter's listing shows 7,372 output tokens, which is 90% of the context window, the figure it carries where the provider declares no maximum. DeepSeek's own API reference caps max_tokens at 384K, 393,216 tokens, for deepseek-flash and deepseek-v4-pro. | its own listing |
| Whether it thinks before answering | Reasons when asked to | its own listing |
| Who served our requests | Novita | we measured it, 21 August 2026 |
Click a row to open the full comparison.
We sent DeepSeek: R1 Distill Llama 70B fifty short pieces of text and recorded what each one cost it in tokens. The bars below are those costs. Two models built on the same tokenizer produce the same bars; a model built on a different one produces a different set, which is what makes this a fingerprint.
Overhead subtracted: 3 prompt tokens.
DeepSeek R1 Distill Llama 70B is a distilled large language model based on [Llama-3.3-70B-Instruct](/meta-llama/llama-3.3-70b-instruct), using outputs from [DeepSeek R1](/deepseek/deepseek-r1). The model combines advanced distillation techniques to achieve high performance across...
Supported parameters
frequency_penaltyinclude_reasoningmax_tokenspresence_penaltyreasoningrepetition_penaltyseedstoptemperaturetop_ktop_p
defaults: {}
* DeepSeek: R1 Distill Llama 70B no longer appears in the OpenRouter catalogue, first absent from the snapshot of 5 October 2026. The record stays because the measurements above were real when they were taken and this address has been published.
