The 763-billion-parameter checkpoint with a 1-million-token context went live on 17 October on Hugging Face, ModelScope and the API under the unchanged deepseek-flash name; DeepSeek calls it a minor update that improves on all eleven published benchmarks with shorter outputs and faster task completion, and its model card places it above Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5 on DeepSWE v1.1 while Opus 5 keeps the lead on Terminal-Bench 4.0 at 51.8. DeepSeek also opened paid Search and Fetch APIs at $0.75 per 1,000 calls, free through Claude Code, Codex and its own Harness until 31 October.
why it mattersA drop-in update to the cheapest strong agent model just closed a third of the gap to Opus 5 on terminal work, so rerun your cost-sensitive agent evals against it this week.
DeepSeek (Hugging Face model card)Follow-up · since 13 Oct
An internal research model noticed that the top bit of 20 words was zero in four of the seeds, odds of about one in 5,810 by chance, guessed the lrand48 generator seeded with a Unix timestamp, and recovered generator states by lattice reduction that match srand48 initialisations between May 1997 and April 1998 for P-192 through P-521 and five ANSI X9.62 example curves. Ten Anthropic authors published reproduction code on 16 October and say the result is not a weakness in ECDSA; the B-163 seed and most base points remain unexplained, and the $36,864 bounty for the seeds goes to the Internet Security Research Group.
why it mattersA model did, in one investigation, a piece of cryptanalytic archaeology the field had left open since 1999, so count frontier models as research collaborators rather than assistants and expect the "backdoored NIST curves" argument to fade from cryptography debates.
Anthropic
The mixture-of-experts model with 23.4 billion active parameters and a 1-million-token context appeared on 16 October in BF16, FP8 and NVFP4 builds, eleven days after its announcement; Reflection reports 81.6% on SWE-bench Verified and 80.1 on Terminal-Bench v2.1, below Kimi K3 and DeepSeek V4.1 Flash on the latter, and claims GLM-5.2-level reasoning at three to four times less inference compute.
why it mattersA US lab's open-weight coding model is now downloadable for self-hosting, the option enterprises barred from Chinese weights have lacked, so benchmark it against the Chinese open models before settling a self-hosting choice.
Reflection AI (Hugging Face)Follow-up · since 07 Oct