DeepSeek launching v4.1 flash cheaper and more capable than v4 pro

Hacker News Discussion ↗

DeepSeek is rolling out its V4.1 Flash large language model, which early users say is faster, cheaper, and often more capable than the existing V4 Pro, especially for coding and agent-style automation tasks. The company is also cutting per-token prices and temporarily routing all V4 Pro API calls to V4.1 Flash at Flash rates, a move some praise as user-friendly and others criticize as risky for production workloads that depend on stable model behavior. Commenters highlight strong value versus Western providers, note issues such as language-mixing and overlong “reasoning” traces, and emphasize that open weights and self-hosting partly mitigate concerns about data privacy and model changes.

Model capabilities and positioning

  • V4.1 Flash is announced as cheaper and more capable than the current V4 Pro; some see this “Flash leapfrogs Pro” pattern as the new normal.
  • Early testers report it as comparable or better than GLM 5.3 Flash and close to frontier models for many coding/agent tasks, though not necessarily top-tier in pure reasoning.
  • Vision support exists in a preview API model, and people are already using it for refactoring kernels, porting code, and other heavy coding tasks.

Pricing and economics

  • Official pricing (off-peak) for Flash is going down across cache hits, cache misses, and output tokens; some confusion arose from comparing different providers and old prices.
  • Peak hours are 2× off-peak, but this was already true; users note it’s still extremely cheap vs US labs, making new automation use cases viable.
  • Several argue the key shift is not “Flash replaces Pro” but “Flash makes many small, repetitive tasks worth automating.”

Performance, tooling, and local deployment

  • Preview tests report 300–400 tok/s, sometimes faster than tools being called.
  • Open weights and GGUF quantizations allow running large Flash/vision models locally on 128 GB RAM via llama.cpp, dwarfstar, and other runtimes.
  • Flash is widely used as a “worker” with more capable planners (e.g., Opus, Sol, GLM Max), especially for code implementation.

Reliability, reasoning, and ergonomics

  • Experiences diverge: some find V4 Flash extremely reliable for coding and agents; others see infinite loops, hallucinations, invalid tool calls, and pathological “but wait” chains, especially via third‑party providers or heavy quantization.
  • Tool-calling robustness varies; harness design (retries, forgiving schemas, structured outputs) is emphasized as critical.
  • Reasoning levels (low/high/max) are criticized: low ≈ off, high ≈ max, causing slow, verbose, and costly runs; people want a true “medium.”

Language behavior and web UI issues

  • Web chat frequently answers or “thinks” in Chinese even when prompted in English, especially after web search; adding “in English” only partially helps.
  • API usage reportedly does not show this issue as often; speculation includes Chinese system prompts and search results overwhelming language instructions.

Production use, model stability, and routing changes

  • DeepSeek will temporarily route all Pro requests to V4.1 Flash at Flash prices.
    • Supporters call this user-friendly and compute-efficient.
    • Critics worry about breaking tuned workflows and want explicit, version-pinned models or “Auto” vs “exact model” options.
  • Broader debate:
    • One side: cloud LLMs are inherently non-deterministic, so pinning exact models is unrealistic; serious users should self-host.
    • Other side: even with nondeterminism, models have characteristic “flavors,” and unannounced swaps undermine evals, safety reviews, and monitoring.

Privacy, jurisdiction, and data use

  • Some remain wary of Chinese providers; others note US/Meta/OpenAI also surveil or reuse data, so jurisdiction may not meaningfully change risk.
  • Open weights let users avoid Chinese servers, though “zero data retention” policies are viewed skeptically industry‑wide.
  • One commenter flags that the official DeepSeek API trains on user data, making self-hosting or non-Chinese proxies attractive for sensitive workloads.