DeepSeek

Open-source flagship V4: 1M context, the #1 open-source score on SWE-bench, at a fraction of what closed-source flagships charge.

DeepSeek is the flagship brand of the open-source model camp — ever since R1 took the world by storm in January 2025 (topping the US App Store within a week of release, and wiping $589 billion off NVIDIA’s market cap in a single day), the name has been shorthand for “open-source catching up with closed-source.” Its current flagship V4 (April 2026 preview) comes in two tiers: V4-Pro (1.6T total parameters / 49B activated) and V4-Flash (284B / 13B activated). The entire lineup defaults to a 1M-token context, ships under the MIT license, and publishes weights on Hugging Face. The hard numbers are brutal: V4-Pro scores 80.6% on SWE-bench Verified — first among open weights and on par with Gemini 3.1 Pro — and as of July 2026 it also leads the overall open-source leaderboard by a clear margin over GLM, Qwen, and Kimi.

Cost remains the killer edge: during off-peak API hours, V4-Pro runs ¥3 per million tokens for input and ¥6 for output (cache-hit input as low as ¥0.025). Measured by output price, that’s roughly 1/29 of Claude Opus 4.8 and 1/35 of GPT-5.5. The local deployment ecosystem is the most mature, too: community GGUF quantizations of V4-Flash landed within 48 hours, and the Q4 quant runs on a single 80GB GPU.

In terms of ecosystem positioning, DeepSeek and Kimi represent two distinct paths for Chinese open-source: DeepSeek bets on “pure models + extreme cost-efficiency.” Coding and reasoning are its strengths, while factual world knowledge still trails Gemini 3.1 Pro. Its API speaks both OpenAI and Anthropic protocols, making migration costs virtually zero.

Note: V4 still carries the preview tag (officially cleared for production use, with the stable release due mid-July). After GA, peak/off-peak pricing kicks in — during Beijing daytime peak hours, the output price doubles to ¥12/M. The legacy model aliases deepseek-chat / deepseek-reasoner will be fully retired on July 24, and existing integrations must migrate to the new model names.

Model Iteration Timeline

Timeline	Version	Highlights
2023-11	DeepSeek Coder	First open-source coding model, where it all began
2024-05	V2	236B MoE, established the low-cost, high-efficiency path
2024-12	V3	671B MoE, extremely low training cost, squaring off against closed-source flagships
2025-01	R1	Reasoning model that took the world by storm; NVIDIA shed $589B in market cap in a single day
2025-08	V3.1	Hybrid thinking/non-thinking in one model, reasoning folded into the mainline
2025-12	V3.2	DSA sparse attention drove major cost reductions
2026-04	V4 preview	Dual Pro/Flash tiers, 1M context, tops open-source coding

Source: Agentopia · Full site model catalog at /llms.txt