DeepSeek

A case of pushing the diagnosis that the bottleneck for agent workloads has shifted from compute to KV-cache storage and bandwidth through the entire model design. Using different active parameters for prefill and decode

At 27% of the per-token FLOPs and 10% of the KV cache of V3.2, it made 1M context something supported all the time

Re-entered the frontier while keeping MIT

Open-weight reasoning models; DeepSeek claimed V3.2-Speciale reached gold-level results at IMO, IOI and other 2025 competitions

97% OCR precision at compression below 10×, about 60% at 20×. A shift in thinking: images as a medium for compressing context

Puts fine-grained sparse attention into a commercial API and cuts long-context costs

The first frontier LLM to appear in a journal after formal peer review. A crack in the practice of relying on corporate technical reports

Billed by DeepSeek as its first step toward the agent era; the API's deepseek-chat and deepseek-reasoner became the non-thinking and thinking modes of one model, merging what had been separate V3 and R1 offerings

The first primary source on the unit economics of inference. DeepSeek itself cautioned that this is an upper bound assuming all usage is paid, and that actual revenue is much lower

Released, one per day over 5 days, working proof of large-scale EP, PD disaggregation and FP8

Moved sparsity from the inference stage to the pretraining stage

Proved that reasoning ability is not a secret recipe of closed labs; self-checking, re-examination and strategy switching emerged. The same report also published the RLVR recipe and a recipe for transferring the capabili

Reasoning models were immediately reproduced as open weights, and 2 Chinese labs declared their pursuit of the frontier on the same day

Established low-precision pretraining at frontier scale

Drew attention with its claim of extreme training-cost efficiency. The model weights themselves moved to MIT only with DeepSeek-V3-0324 in March 2025

Productized the prefix cache; an internal serving optimization surfaced in the API price list

An open-weight code model that DeepSeek claimed matched GPT-4 Turbo on code-specific tasks

A third path, distinct from MQA and GQA, that compresses KV into a low-dimensional latent

Group relative policy optimization without a value function later became the de facto standard for RL in reasoning models

A general-purpose DeepSeek model released down to its intermediate checkpoints; DeepSeek claimed the 67B Base outperformed Llama 2 70B Base in reasoning, coding, math and Chinese comprehension