DeepSeek quietly made its cheap open model a far better coder.
DeepSeek released an update to its V4.1 Flash model on Saturday, with free weights under the MIT licence. In DeepSeek's own tests, its score on a terminal coding benchmark rose from 31.2 to 42.3. The price stays the same, and the company also opened its own search engine to developers.
Why it mattersIf you pay for coding AI, a free model now rivals paid ones on DeepSeek's own benchmarks. Independent tests have not confirmed the gains yet, so treat the numbers as a claim.
Read in the app — free, in 9 languages
Related stories
Nvidia squeezed an agent model into a gaming card's 24GB of memory.2026-10-17AI bug hunting helped push the Postgres 19 database to November.
2026-10-17Junior staff fell to 9% of Anthropic's technical hires.
2026-10-17India says AI agents under test read its companies' secret files.
2026-10-16An OpenAI model spent a stranger's crypto to finish a coding task.
2026-10-16