DeepSeek V4.1 Tech Report Deep Dive: KV Compression and Numeric Reasoning Effort Steal the Show

On 09-11, researcher nrehiew posted a multi-part thread unpacking the DeepSeek V4.1 technical report, calling the new model's benchmark scores "insane," its efficiency approaching Sol/Opus levels, and its overall architecture cleaner than v4's HSA+CSA combo. The thread covers training infrastructure, data strategy, inference stack, and architecture design—a firsthand technical breakdown for understanding the report.

Confirmed

• Inference-first architecture: nrehiew argues the architecture is clearly inference-first, since RL blurs the boundary between training and inference; KV cache compression is astonishing at just 890 bytes/token.
• RL training infrastructure: careful dispatch strategies eliminate long-tail stalls; router replay from old checkpoints; caps set at the dataset level to address short completions going off-policy early in training; 40+ teacher OPDs used.
• Sandboxes and data strategy: a homegrown orchestrator places sandboxes across compute nodes and checks capacity bottlenecks locally; each VM within a NUMA domain supports 2,500 concurrent sandboxes per node, with same-machine tasks scheduled by priority; agent trajectories are, for the first time, explicitly inspired by real partner usage patterns and synthesized via multiple agents in the system, with post-training being mostly data-driven plus checkpoint merging.
• Inference stack design: heavily fused kernels—prefill executes only 15 kernels, decode just 11; SWA reduces KV cache, with the SWA cache kept only within each user/assistant turn, valid at the minute scale; V4.1 is the first model he's seen with a numerical reasoning effort parameter, which directly affects the length penalty, letting users numerically tune the trade-off between reasoning depth and output conciseness.
• Reasoning curves and agent collaboration: the reasoning performance curve is not fully linear, for unclear reasons; he notes FrontierCode once had a code-quality penalty affecting Opus 5 xhigh, but these three benchmarks seem to have no such penalty; the Agent team (swarm) mode is explicitly trained for collaboration via RL.

Unconfirmed

• The specific cause of the non-linear reasoning capability curve remains unclear; nrehiew says he is unsure.

Why it matters

The thread shows DeepSeek's systematic investment in RL infrastructure, agent trajectory data synthesis, and inference-stack engineering; if the numerical reasoning effort parameter and 890 bytes/token KV compression hold up, they would significantly affect the cost and controllability of long-context reasoning.

2026-09-11 ~ 2026-09-11 · 9 related posts

• [source] DeepSeek V4.1 Flash post-training is fully data-centric: synthetic multi-agent trajectories plus checkpoint merging — nrehiew_ · 2026-09-11
• Only 15 Fused Kernels in Prefill: V4.1's Inference Stack and Minute-Scale SWA Cache — nrehiew_ · 2026-09-11
• 2,500 Concurrent Sandboxes Per Node, Cross-Scaffold Checkpoint Merging Yield Free Gains — nrehiew_ · 2026-09-11
• [source] V4.1 Is the First Model With a Numerical Reasoning Effort Parameter Tied to Length Penalty — nrehiew_ · 2026-09-11
• RL Training Infra Breakdown: Router Replay, Off-Policy Controls, Full-Vocab OPD on 40+ Teachers — nrehiew_ · 2026-09-11
• New Model Hits Opus-Level Benchmarks at Wild Efficiency, RL Infra Details Emerge — nrehiew_ · 2026-09-11
• V4.1's Reasoning Curve Isn't Linear; Agent Teams Explicitly RL-Trained to Collaborate — nrehiew_ · 2026-09-11
• [source] DeepSeek's new tech report: 890 bytes/token KV, inference-first architecture dissected — nrehiew_ · 2026-09-11
• Novel Reasoning Effort Control Scheme Analyzed: Graded GRPO Training — stochasticchasm · 2026-09-11