DeepSeek-V4.1-Flash-1017

AFFORDABLE CLASS: sub-8B-active MoE on the prefill side. The 1017 card does not restate parameter counts; it calls this a minor update of DeepSeek-V4.1-Flash, and the two repos' config.json text_config blocks are byte-identical (fetched 2026-10-17). The parent card states 552B backbone, 8B activated per token during prefill and 16B during decode, from a Causal Encoder-Decoder split (20-layer causal encoder, 20-layer decoder) where the decoder's global KV is projected from the final encoder hidden states rather than recomputed per decoder layer. Decode activation at 16B is above the strict sub-8B bar; prefill at 8B is exactly on it, and the design target is input-heavy agent workloads, which is the consumer-runnable class this record is filed under. No tok/s is published on either card. PARAM COUNTING: HF safetensors total is 763,205,315,794 parameters (BF16 1.976B, F32 42.3M, F8_E4M3 204.0B, I8 557.2B). params_b is recorded as the card's 552 backbone, not 763. The difference is the Engram conditional-memory tables: engram_num_embeddings [384006168, 384016682] at engram_n_heads 8 x engram_head_dim 256, stored as I8, which the parent card separates out as 196B sparsely accessed parameters. 552 + 196 = 748, close to but not equal to the 763B safetensors total; the residual is the vision encoder, MTP/DSpark heads, and unembedding, so the 552 figure is the vendor backbone claim and 763.2 is the file total. KV: kv_cache_bytes_per_token is the parent card's stated global KV footprint of 890 bytes/token (vendor_claim), not a derivation. The card attributes it to CSA2 (per-layer Full / Reindex / Reuse modes sharing main KV and indexer K) plus FP4 main KV (E2M1, one E4M3 scale per 16 channels), about 1/4 of DeepSeek-V4-Flash. config has no kv_lora_rank, so the V2/V3 MLA formula does not apply. compress_ratios over 40 layers is [0, 0] + eighteen 2s + twenty 1s + three 0s; sliding_window 128; kv_source_layer_ids [2, 8, 14, 20]; index_source_layer_ids [2, 8, 14, 20, 24, 28, 32, 36]; index_n_heads 32, index_head_dim 128, index_topk 512; candidate pool from layer 20 (candidate_topk_blocks 2048, candidate_block_size 8). SWA Bounded Replay reconstructs sliding-window KV from the last n_win tokens so SWA KV is not persisted, cutting the persistent KV footprint to roughly 1/8 of V4-Flash. An exact byte derivation needs modeling_deepseek_v41, which this repo does not ship (inference/model.py is the reference implementation). ARCHITECTURE from config.json: model_type deepseek_v41, hidden 5120, 40 layers, 64 heads / 1 kv head / head_dim 512, qk_rope_head_dim 64, q_lora_rank 1280, o_lora_rank 1024, o_groups 8, silu with swiglu_limit 10. 384 routed experts + 1 shared, 6 routed per token, moe_intermediate_size 2304, scoring_func sqrtsoftplus, topk_method noaux_tc, routed_scaling_factor 1.5. Context 1,048,576 via YaRN factor 16 over original_max_position_embeddings 65536, rope_theta 10000. Engram on layers [1, 14], max ngram 4, engram vocab 16,000,000. DSpark speculative decoding: num_nextn_predict_layers 3, dspark_target_layer_ids [37, 38, 39], dspark_n_routed_experts 128, dspark_num_experts_per_tok 3, dspark_markov_rank 256. Vision: DeepSeek-ViT, 32 layers, hidden 1024, 16 heads, intermediate 2816, patch 14, 3x downsample, max_image_tokens 1024. Weights ship mixed: fp8 e4m3 dynamic with 32x32 blocks, expert weights fp4. flops_per_token_active = 2 x 8e9 = 1.6e10 uses the prefill activation; decode is 2 x 16e9 = 3.2e10, so the field understates decode. BENCHMARKS are vendor_claim on the 1017 card (max reasoning effort implied by the parent methodology, temperature 1.0, top_p 1.0 for this checkpoint's recommended sampling vs 0.95 on the parent table): HLE without tools 42.6 (44.0 text-only) vs parent 36.8 (39.1), Terminal-Bench 4.0 42.3 vs 31.2, DeepSWE v1.1 77.0 vs 74.2, ProgramBench 29.1 vs 20.3, NL2Repo-Bench 68.2 vs 65.4, Automation-Bench v1.0.6 60.5 vs 54.8, Agents' Last Exam 35.4 vs 31.8, Apex Bench v1.1 72.1 vs 64.6, Chartography with tools 83.0 vs 78.9, BabyVision with tools 92.5 vs 89.6, ZeroBench-main with tools 51.0 vs 49.0. MIT. Released 2026-10-17 (HF createdAt 2026-10-17T02:40:38Z).