DeepSeek swapped its Flash model overnight. I pulled the new config and did the cache math.

Описание и применение

DeepSeek внезапно заменил модель под тем же именем API, что привело к непредсказуемому изменению поведения. Пользователи могут скачать новый конфиг, посчитать размер кэша и закрепить конкретный чекпоинт, чтобы избежать подобных сбоев.

Задача

A silent model swap under the same API name caused users to unknowingly interact with a different checkpoint, breaking reproducibility and evaluation consistency.

Результат

The new checkpoint is larger (≈510 GB) and offers better benchmarks, but its per‑token cache cost varies with context length; the advertised 890 bytes/token is a lower bound. The silent swap highlights the risk of relying on vendor aliases without version pinning.

Рабочий процесс

• 1. Download the new config.json from Hugging Face. 2. Extract key parameters (num_hidden_layers, compress_ratios, kv_source_layer_ids, etc.). 3. Sum the safetensor shard sizes to compute total model size. 4. Perform cache‑size calculations using the provided formulas. 5. Compare calculated per‑token cache bytes to the 890 bytes figure. 6. Review DeepSeek’s pricing to understand cache read cost. 7. Pin the checkpoint name (DeepSeek‑V4.1‑Flash‑1017) in deployments to avoid silent swaps.

Доказательства

On October 17, DeepSeek replaced the model behind its deepseek‑flash API endpoint.

Источник

The new checkpoint is DeepSeek‑V4.1‑Flash‑1017, on Hugging Face under MIT, and it was the top story on Hacker News this morning (903 points when I checked).

Источник

The circulating 890 figure sits below even my most aggressive bound, which means quantization below bf16 and the sparse index (index_topk: 512) are doing the rest.

Источник