inference/research
verifiedhigh · 0.88updated 2026-10-17

deepseek-ai/DeepSeek-V4.1-Flash-1017 model card and config.json

repo

Fields

id
sources/hf-deepseek-v4-1-flash-1017
Publisher
DeepSeek
Kind
repo
Published
2026-10-17
Accessed
2026-10-17
notes
Primary source for the DeepSeek-V4.1-Flash-1017 record. HF API createdAt 2026-10-17T02:40:38Z, lastModified 2026-10-17T05:46:55Z, sha 1bae0e1088993b9b82de08f59de5be5a34e42170. Card (raw README fetched 2026-10-17) calls it a minor update to DeepSeek-V4.1-Flash that improves reasoning, coding, agentic, and visual benchmarks while reducing task execution time and output length; it does not restate parameter counts. Architecture and the 8B-prefill / 16B-decode activation figures are inherited from the parent card deepseek-ai/DeepSeek-V4.1-Flash (same fetch window), whose config.json text_config is byte-identical to this repo's config.json text_config: model_type deepseek_v41, architectures DeepseekV41ForCausalLM, hidden_size 5120, num_hidden_layers 40, num_attention_heads 64, num_key_value_heads 1, head_dim 512, qk_rope_head_dim 64, q_lora_rank 1280, o_lora_rank 1024, n_routed_experts 384, n_shared_experts 1, num_experts_per_tok 6, moe_intermediate_size 2304, max_position_embeddings 1048576, rope_scaling yarn factor 16 over original 65536, sliding_window 128, compress_ratios (2 full, 18 ratio-2, 20 ratio-1, 3 ratio-0), engram_layer_ids [1, 14], engram_num_embeddings [384006168, 384016682], num_nextn_predict_layers 3, dspark draft experts 128 / 3 per token, vocab 129280. safetensors.parameters total 763,205,315,794 (BF16 1,976,441,856; F32 42,307,282; F8_E4M3 204,015,223,296; I8 557,171,343,360) across 48 shards, usedStorage 510,296,920,807 bytes. The 763B figure counts the Engram embedding tables (I8), not the 552B backbone the parent card advertises. quantization_config: fp8 dynamic, weight_block_size [32, 32], expert_dtype fp4. License MIT. Benchmark table on the 1017 card is a vendor claim (HLE, Terminal-Bench 4.0, DeepSWE v1.1, ProgramBench, NL2Repo, Automation-Bench, Agents' Last Exam, Apex Bench, Chartography, BabyVision, ZeroBench). The card gives no tok/s. Parent technical report is DeepSeek_V41_Tech_Report.pdf on the V4.1-Flash repo, not a separate arXiv id.

Notes

Primary source for the DeepSeek-V4.1-Flash-1017 record. HF API createdAt 2026-10-17T02:40:38Z, lastModified 2026-10-17T05:46:55Z, sha 1bae0e1088993b9b82de08f59de5be5a34e42170. Card (raw README fetched 2026-10-17) calls it a minor update to DeepSeek-V4.1-Flash that improves reasoning, coding, agentic, and visual benchmarks while reducing task execution time and output length; it does not restate parameter counts. Architecture and the 8B-prefill / 16B-decode activation figures are inherited from the parent card deepseek-ai/DeepSeek-V4.1-Flash (same fetch window), whose config.json text_config is byte-identical to this repo's config.json text_config: model_type deepseek_v41, architectures DeepseekV41ForCausalLM, hidden_size 5120, num_hidden_layers 40, num_attention_heads 64, num_key_value_heads 1, head_dim 512, qk_rope_head_dim 64, q_lora_rank 1280, o_lora_rank 1024, n_routed_experts 384, n_shared_experts 1, num_experts_per_tok 6, moe_intermediate_size 2304, max_position_embeddings 1048576, rope_scaling yarn factor 16 over original 65536, sliding_window 128, compress_ratios (2 full, 18 ratio-2, 20 ratio-1, 3 ratio-0), engram_layer_ids [1, 14], engram_num_embeddings [384006168, 384016682], num_nextn_predict_layers 3, dspark draft experts 128 / 3 per token, vocab 129280. safetensors.parameters total 763,205,315,794 (BF16 1,976,441,856; F32 42,307,282; F8_E4M3 204,015,223,296; I8 557,171,343,360) across 48 shards, usedStorage 510,296,920,807 bytes. The 763B figure counts the Engram embedding tables (I8), not the 552B backbone the parent card advertises. quantization_config: fp8 dynamic, weight_block_size [32, 32], expert_dtype fp4. License MIT. Benchmark table on the 1017 card is a vendor claim (HLE, Terminal-Bench 4.0, DeepSWE v1.1, ProgramBench, NL2Repo, Automation-Bench, Agents' Last Exam, Apex Bench, Chartography, BabyVision, ZeroBench). The card gives no tok/s. Parent technical report is DeepSeek_V41_Tech_Report.pdf on the V4.1-Flash repo, not a separate arXiv id.

Cross-references

This record references no other records.