DeepSeek Drops 60-85% Inference Speed Gain, Eyes Fundraise
DeepSeek open-sourced inference optimizations that cut generation time by up to 85%, while its founder reportedly started fundraising after watching Anthropic's Mythos preview.
What's happening
DeepSeek published open-source inference optimizations this week, claiming 60 to 85 percent faster generation โ a technical release, not a benchmark press release. Separately, reporting suggests founder Liang Wenfeng began fundraising in April after Anthropic previewed its Mythos model, implying the competitive pressure of frontier lab releases is now moving DeepSeek toward outside capital. The company also launched DeepSeek V4 Pro, featuring a one-million-token context window positioned at agent coding and long-session engineering work.
The numbers
Velocity 4.9 on thin recent volume โ just 12 signals โ across 4 sources, nearly all technical or developer-facing. Breadth is low: this conversation is happening in AI/engineering circles, not mainstream finance or retail.
Our read
A 60-85% inference speedup is a number that matters if it holds under scrutiny โ open-source release means the community will verify it quickly. The fundraising signal is the more strategically interesting detail: DeepSeek has operated lean on Chinese capital; moving toward outside funding after a competitor's model preview suggests they see a capability gap forming that requires a spend-up. Low breadth here is a feature, not a bug โ watch for this to widen when the inference benchmarks start getting replicated publicly.
Sources we checked
- DeepSeek open-sources inference optimizations with 60โ85% faster generation [pdf] โ github.com
- esengine/DeepSeek-Reasonix โ github.com
- RT by @zephyr_z9: official dsv4 spec dec and draft model @teortaxesTex github: https://github.com/deepseek-ai/DeepSpec huggingface: https://huggingface.co/deeps โ @nitter.privacyredirect.com.zephyr.z9@rss-parrot.net
- Verdent (@verdent_ai)DeepSeek V4 Pro๊ฐ Verdent์์ ์ง์๋ฉ๋๋ค. ์์ด์ ํธ ์ฝ๋ฉ, ๊ณ ๊ธ ์ํ, ๋ณต์กํ ์ถ๋ก ์ ๋ง์ถ ๋ชจ๋ธ๋ก, 100๋ง ํ ํฐ ์ปจํ ์คํธ์ ๋์ฉ๋ ์ฝ๋๋ฒ ์ด์ค/๊ธด ์์ง๋์ด๋ง ์ธ์ ์ ์ํ ์ฅ๋ฌธ ์ปจํ ์คํธ ์ฒ๋ฆฌ๊ฐ ํต์ฌ์ ๋๋ค. ์ค๋ฌด ์ฝ๋ ์์ ์ฉ ๋ชจ๋ธ ์ง์ โ @sayzard@mastodon.sayzard.org
- Elestio (@elest_io)์คํ์จ์ดํธ ๋ชจ๋ธ์ด ๋๋ถ๋ถ์ ์ค์ ๋น๋์์๋ ์ด๋ฏธ ์น๋ฆฌํ๋ค๋ ์ฃผ์ฅ์ ๋๋ค. ์์ฒด ํ๋์จ์ด์์ ์ด์ํ๊ณ , ์์ฒด ๋ฐ์ดํฐ๋ก ํ์ธํ๋ํ๋ฉฐ, ํ ํฐ๋น ๊ณผ๊ธ ๋์ ์ ๊ธฐ๋ฃ๋ฅผ ์ง๋ถํ๋ ๋ฐฉ์์ด ๋ ์ ๋ฆฌํ๋ค๋ ๊ด์ ์ ๋๋ค. Alibaba Qwen, DeepSeek, Z.ai, โ @sayzard@mastodon.sayzard.org
- Max For AI (@MaxForAI)The Information ๋ณด๋์ ๋ฐ๋ฅด๋ฉด DeepSeek ์ฐฝ์ ์ Liang Wenfeng์ด ์ฌํด 4์ Anthropic์ Mythos ํ๋ฆฌ๋ทฐ ๊ณต๊ฐ๋ฅผ ๊ณ๊ธฐ๋ก ์๊ธ ์กฐ๋ฌ์ ์์ํ ๊ฒ์ผ๋ก ๋ณด์ ๋๋ค. ๊ฒฝ์์ฌ ๋ชจ๋ธ/์ ํ ๋ฐํ๊ฐ ์ง์ ์ ์ผ๋ก ํ์ ํฌ์ ํ๋จ์ โ @sayzard@mastodon.sayzard.org
- DeepSeek lanรงa recrutamento em massa para contratar novos profissionais e acelerar a chegada da sua tecnologia de IA ao mercado. A empresa chinesa busca duplica โ @tugatech@masto.pt
- got DeepSeek to get the text from the image. Then fed the text to GoogleGemini and told it to create calendar appointments. It did it, despite telling me this w โ @andyjennings@aus.social
See today's live board
This story is frozen as published on June 28, 2026. The live board shows what is moving right now, scored against real money.
Open this story live โ See today's live boardMore AI stories
- Kimi K3 Tops Coding Benchmarks and Pulls Down Chip Stocks July 17, 2026
- China's Kimi K3 Tops Code Benchmark, Rattles Chipmakers July 17, 2026
- Claude Sonnet 5 Reaches Near-Opus Benchmarks at Lower Cost June 30, 2026
- US Clears Anthropic Mythos for 100-Plus Organizations June 27, 2026
- US Government Will Individually Approve Who Gets GPT-5.6 June 27, 2026
Research, not financial advice. Sort Brick reports what its data showed at the time of publication. It is not a recommendation to buy or sell anything โ you decide.