DeepSeek posts V4.1-Flash-1017 weights on Hugging Face under MIT, and launches Search and Fetch APIs at $0.75 per 1,000 successful requests

Start with the links and the facts. The Hugging Face repo is deepseek-ai/DeepSeek-V4.1-Flash-1017, its page lists License: mit, the files ship as safetensors, and the tags include Image-Text-to-Text, transformers, safetensors, fp8 and 8-bit precision. The repo's config.json is specific: the architecture field is DeepseekV41ForCausalLM, quantization_config sets quant_method to fp8 with a dynamic activation scheme, weight_block_size of [32,32], scale_fmt of ue8m0 and expert_dtype of fp4. Dense weights run in FP8, expert weights are squeezed further to FP4.

The architecture parameters are in the public config too, so nothing here needs guessing. In text_config, num_hidden_layers is 40, hidden_size 5120, vocab_size 129280, num_attention_heads 64 against num_key_value_heads 1, head_dim 512. The MoE block declares 384 routed experts plus one shared expert with num_experts_per_tok at 6. Context fields are max_position_embeddings 1048576, rope_scaling of type yarn with factor 16, and sliding_window 128, plus 3 num_nextn_predict_layers for multi-token prediction. On the vision side, vision_config lists 32 layers, hidden 1024, patch 14 and max_image_tokens 1024.

The interface changes come in three parts, all documented officially. First, the model name: the Models & Pricing page says to use deepseek-flash, model version DeepSeek-V4.1-Flash, 1M context, 384K max output, thinking mode on by default and switchable, with JSON output, tool calls, the Responses API and Anthropic-format support, while vision works on Flash but not on V4 Pro. A footnote is explicit that the legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the models behind them are retired and their requests are served by V4.1-Flash and billed at the Flash price.

Second is the price table, third is the new tool API. Prices are per million tokens: cache-hit input is $0.003 off-peak and $0.006 at peak; cache-miss input is $0.15 off-peak and $0.3 at peak; output is $0.6 off-peak and $1.2 at peak. Off-peak is half of peak, with peak defined as 01:00-04:00 and 06:00-10:00 UTC Monday to Friday excluding Chinese public holidays, everything else including weekends billed off-peak; concurrency is capped at 2500. On tools, the docs add a Search & Fetch section: Search is POST https://api.deepseek.com/v1/tools/search and Fetch is POST https://api.deepseek.com/v1/tools/fetch, both using the same Bearer auth and billed per successful request at $0.75 per 1,000 each, with a QPS limit of 5 each, and failures, empty results or timeouts not billed. There is also a timing note: search requests issued through Claude Code's built-in Web Search or Codex's web_search feature are not charged during the first two weeks after the Search API launch.

One detail is easy to get wrong: the API release date and the weights repo date are not the same day. The official Change Log dates the DeepSeek-V4.1-Flash release to September 10, 2026, describing it as the smallest model in the new architecture family with native multimodal visual understanding, while the Hugging Face repo carries a 1017 suffix, a snapshot from October 17. Framing this as a brand-new model launched today would be wrong; what happened today is the open weights plus interface additions, namely Search/Fetch and the pricing page update.

The boundaries deserve the same clarity. First, although the weights are MIT, the Hugging Face model card is minimal — a row of badges and a link to a technical report — with no parameter count, training-data description, benchmark scores or safety evaluation, so this article deliberately avoids 'X billion parameters' or 'score Y' claims: config.json proves layer counts, expert counts and context settings, not capability. Second, expert weights are FP4, so real memory use and accuracy depend on how well a given inference stack supports FP4 experts, and the official pages publish no VRAM or throughput figures. Third, the $0.75 Search and Fetch price is per successful request and independent of result count; count only controls how many results come back, so do not model it per token. Fourth, Search is limited to 5 QPS per account and Fetch adds a per-domain limit on top, so size any crawling job against those limits before hitting 429s.