DeepSeek V4.1-Flash-1017 Is the Least Hallucinatory DeepSeek Model, but Pareto-Dominated by Sol
teortaxesTex · x · 2026-10-19
teortaxesTex analyzes DeepSeek V4.1-Flash-1017 on the Artificial Analysis leaderboard and questions its cost accounting.
Key findings:
- Hallucination: V4.1-Flash-1017 is the least hallucinatory DeepSeek model of 2026 (45-46 would be decent, better than Grok), though it regresses on knowledge and LCR vs its predecessor
- Value: It holds a spot on the speed Pareto frontier, but at peak price it's Pareto-mogged by Sol
- Cost skepticism: The model is as verbose on AA as V4.1F, so the claimed cost savings vs V4.1 Flash don't add up — the author asks Artificial Analysis whether they mistakenly priced 1017 at off-peak rates
Charts show V4.1 Flash (max) at $0.49/task with 250M output tokens (verbosity rank 47/47), while the off-peak version scores 45 at $0.13/task, compared to Sol 6.1 Low (42, $0.13, 1.2 min/task) and MiMo-V2.6 Pro (46, $0.13, 2 min/task).
Related event: DeepSeek V4.1-Flash hits lowest hallucination rate but regresses elsewhere(2 posts)→
More from Infra
- Owning vs renting vs API for inference comes down to GPU utilization — Milan_Slov26 · 2026-10-19
- Anthropic signed 57% of this year's $157B in neocloud deals, dwarfing OpenAI — AccBalanced · 2026-10-19
- Open decision model D3 tops Jev Index, ships vLLM recipes from 0.6B to 27B — AccBalanced · 2026-10-19
- 2-GPU R9700 vLLM rig hard power-offs: tensor-parallel all-reduce transient, not wattage, kills 1200W PSUs — brainchillzZ · 2026-10-19
- AI performance engineering meetup: GPU rooflines, agent trace-to-training loop, Homa transport — cfregly · 2026-10-19
- Ultra Ethernet Spec 1.1 Lands With Scale-Up Networking and In-Network Computing — thoefler · 2026-10-19