DeepSeek V4.1 Flash 1017 (Max) — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- DeepSeek released V4.1 Flash 1017 (Max) on October 17, 2026 — a Mixture-of-Experts model with 552B total parameters (16B active), open weights, and an MIT license permitting self-hosting
- The model scores 45 on the Artificial Analysis Intelligence Index, well above the 19 median for similarly sized open-weight peers
- Output runs at 230.4 tokens per second, nearly 3x the class median of 77.9 t/s, with a 1.08-second time-to-first-token versus a 1.69s median
- Pricing lands at $0.30 per 1M input tokens and $1.20 per 1M output tokens, both better than the class medians of $0.43 and $1.68
- The model ships with a 1M-token context window and multimodal support (text and image input, text output)
- As a reasoning model using chain-of-thought, it generates 210M output tokens per Intelligence Index run — notably verbose against the 140M median
- Available through 5 API providers, giving buyers multiple routing options
Why it matters: DeepSeek just delivered a model that pairs above-average intelligence (45 vs 19 median), top-class throughput (3x the median tokens/second), and below-median pricing — all under an MIT license with weights available through 5 API providers. For buyers weighing open-weight alternatives to proprietary frontier models, that combination is rare, though the model's 210M-token verbosity per benchmark run (vs 140M median) inflates spend on reasoning-heavy workloads.
Ask SkimNews
