PublishersDeepSeek

DeepSeek

15 documents read from this publisher.

deepseek-ai/DeepSeek-V4-Flash-Vision-Exp · Hugging Face

first party release · model card · 2026-09-10 · 15 techniques

Mixture of Experts FP8 KV-cache quantization DSpark vLLM DeepSeek Harness Minimal mode Hyper-Connections Compact TXT image-path notation Continued training for visual understanding DFlash attention DSpark forward path OpenAI-style JSON content blocks Same-checkpoint target and draft weights Separate encoding and inference modules Vision encoder and multimodal aligner Visual modules for multimodal understanding

deepseek-ai/DeepSeek-V4.1-Flash · Hugging Face

first party release · model card · 2026-09-10 · 22 techniques

DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

technical report · 2026-04-26 · 76 techniques

Multi-Token Prediction DeepSeek Sparse Attention Group Relative Policy Optimization Multi-Teacher On-Policy Distillation FP8 KV-cache quantization Muon Sliding Window Attention On-Policy Distillation Manifold-Constrained Hyper-Connections Quantization-Aware Training Generative Reward Model Shared Experts Compressed Sparse Attention Auxiliary-loss-free load balancing Heavily Compressed Attention Rotary Position Embedding AdamW FP4 Quantization Batch-invariant deterministic kernels Cross-turn persistent reasoning history FP4 Quantization-Aware Training Periodic cache checkpointing RMSNorm Sample-level attention masking SwiGLU SwiGLU clamping Anticipatory Routing DeepSeekMoE Filtering batched auto-generated and templated content Full-vocabulary logit distillation Multi-query attention Nesterov momentum QK-Clip Quick Instruction tokens XML-based tool-call schema with DSML token Zero SWA caching Agentic search All-to-all Gradient Exchange with Local FP32 Summation Attention sink Avoiding split-K

deepseek-ai/DeepSeek-V4-Flash-0731 · Hugging Face

first party release · model card · 2026-04-26 · 5 techniques

Configurable reasoning effort Speculative decoding DSpark OpenAI-compatible message encoding without Jinja template recommended sampling parameters for deployment

deepseek-ai/DeepSeek-V4-Flash · Hugging Face

first party release · model card · 2026-04-26 · 9 techniques

Mixture of Experts Hybrid Attention Group Relative Policy Optimization Muon On-Policy Distillation Manifold-Constrained Hyper-Connections Compressed Sparse Attention Heavily Compressed Attention FP4+FP8 mixed precision

deepseek-ai/DeepSeek-V4-Pro · Hugging Face

first party release · model card · 2026-04-26 · 11 techniques

Mixture of Experts Hybrid Attention Group Relative Policy Optimization Supervised fine-tuning Muon On-Policy Distillation Manifold-Constrained Hyper-Connections Compressed Sparse Attention Heavily Compressed Attention FP4+FP8 mixed precision Post-training pipeline

deepseek-ai/DeepSeek-V4-Pro-0813 · Hugging Face

first party release · model card · 2026-04-26 · 4 techniques

Configurable reasoning effort DSpark OpenAI-compatible message encoding recommended sampling parameters

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

research paper · technical report · 2025-12-02 · 30 techniques

DeepSeek Sparse Attention Group Relative Policy Optimization Discard-all context management Agentic data synthesis pipeline KL alignment loss Lightning Indexer Off-Policy Sequence Masking Specialist Distillation Unbiased KL Estimate Automatic synthesis of task-oriented RL environments Dense Warm-up Stage Discard-75%Fine-grained token selection Keep Sampling Mask Mixed Reinforcement Learning Model-based refusal and inability filtering MQA Mode of Multi-Head Latent Attention Multi-stage candidate filtering Parallel-fewest-step sampling Pass@100 filtering Sparse-attention continued pre-training with joint model and indexer optimization Thinking context management for tool use Trajectory summarization and rollout re-initiation Environment, toolset, task, and solution synthesis pipeline Generate-verify-refine loop Keep Routing Longest-trace selection Prompt-based cold start for tool use Sample test and length-constraint filtering Training agentic policies with real-world tools

deepseek-ai/DeepSeek-V3.2 · Hugging Face

first party release · model card · 2025-12-01 · 4 techniques

DeepSeek Sparse Attention Agentic data synthesis pipeline Scalable reinforcement learning post-training Thinking with tools

DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient | DeepSeek API Docs

first party release · official blog · 2025-09-22 · 7 techniques

Mixture of Experts Causal Encoder-Decoder (CED) architecture Compressed KV caching Reinforcement Learning Post-Training Asymmetric architecture Peak/off-peak pricing Pre-training methods

DeepSeek V4 Preview Release | DeepSeek API Docs

first party release · official blog · 2025-09-22 · 5 techniques

Mixture of Experts DeepSeek Sparse Attention 1M-token context window Thinking mode selection Token-wise compression

Change Log | DeepSeek API Docs

first party release · official blog · 2024-05-17 · 9 techniques

Configurable reasoning effort Function calling Chat Prefix Completion disk-based context caching Fill-in-the-middle (FIM) completion hybrid reasoning architecture JSON mode Native multimodal visual understanding Post-training optimization

deepseek V4 model card EN

first party release · technical report · 9 techniques

Mixture of Experts Multi-Token Prediction Hybrid Attention DeepSeek Sparse Attention Muon Manifold-Constrained Hyper-Connections Compressed Sparse Attention Heavily Compressed Attention DeepSeekMoE

paper

research paper · technical report · 56 techniques

DeepSeek Sparse Attention Group Relative Policy Optimization Default thinking mode Discard-all context management Generative Reward Model KL alignment loss Continual pretraining Lightning Indexer Off-Policy Sequence Masking Specialist Distillation Unbiased KL Estimate Automatic synthesis of task-oriented RL environments Dense Warm-up Stage Discard-75%Domain-specific KL regularization strength Fine-grained token selection Generate-verify-refine loop Keep Sampling Mask Length-adjusted RL reward Mixed Reinforcement Learning Model-based refusal and inability filtering MQA Mode of Multi-Head Latent Attention Multi-stage candidate filtering Parallel-fewest-step sampling Pass@100 filtering Reinforcement learning on synthetic agentic data Sparse-attention continued pre-training with joint model and indexer optimization Thinking context management for tool use Trajectory summarization and rollout re-initiation Coding environments from GitHub issue–PR pairs Context management method Detached indexer-input optimization F2P/P2F acceptance criterion Heterogeneous answer-generation agents Iterative task-difficulty escalation Jupyter Notebook code interpreter Keep Routing Language consistency reward Multi-agent search-agent training-data generation Multi-Head Latent Attention MHA/MQA Modes

DeepSeek V41 Tech Report

technical report · 185 techniques

Mixture of Experts Configurable reasoning effort Speculative decoding Muon DSpark Sliding Window Attention On-Policy Distillation Causal Encoder-Decoder (CED) architecture Shared Experts Compressed Sparse Attention 2 Auxiliary-loss-free load balancing Deployment-time scalar effort control Engram FP4 KV-cache quantization Hierarchical Sparse Indexer Per-Head Muon AdamW SWA Bounded Replay 2D rotary position embedding Offline checkpoint merging Reindex Mode Reuse Mode RMSNorm Sample-level attention masking Single-Pass mHC SwiGLU SwiGLU clamping 3×3 pixel-unshuffle downsampling Automated data synthesis and environment-construction pipelines Best-fit packing Co-located RL training Communication-Computation Overlap Containerized coding-environment construction with self-testing and trace removal CSA2 Full Mode Decoder SWA Bounded Replay Decoupling agent rollout into sandbox and worker container Deterministic pre-splitting of ultra-long documents Effort-dependent exponential token-penalty schedule Encoder-Prefill-Decode disaggregation High-confidence ProgramBench subset