DeepSeek v4.1 Flash

Original Article ↗ Hacker News Discussion ↗

DeepSeek’s new V4.1 Flash model is a massive open-weights multimodal LLM that combines a 552B-parameter sparse backbone with 196B “Engram” memory, using a novel causal encoder–decoder design to slash KV cache size and deliver very high throughput at relatively low API prices. Commenters highlight that while the model is too large for typical local setups without extreme hardware or aggressive quantization, its efficiency, reasoning controls, and strong benchmarks put it near frontier systems like GPT-6 Astra and Claude Opus for many coding, security, and agentic workloads. The release also intensifies broader debates over Chinese vs. US labs, open vs. closed models, prompt data retention, and whether safety and “welfare” work is genuine risk mitigation or mostly marketing and regulatory positioning.

Model architecture & features

  • V4.1 Flash is a major redesign, not a minor point release.
  • Uses a Causal Encoder–Decoder (CED) transformer: ~20-layer causal encoder + 20-layer decoder.
  • Only ~8B params activated per token during prefill and 16B during decode, despite:
    • 552B “backbone” params (mostly ~FP4, ~306GB)
    • 196B FP8 “Engram” / PLE-like memory (~204GB), treated more like a key–value store and storable on SSD.
  • KV cache is heavily compressed: 890–900 bytes per token (1GB for 1M context), about 1/4 of previous V4 Flash.
  • Supports multimodal (vision) and controllable reasoning effort (1–100), mapping to a few internal levels (low/high/max).

Performance & benchmarks

  • Many commenters say it’s extremely fast in API form: often 250–400+ tokens/s, though some see lower speeds via certain OpenRouter providers.
  • Benchmarks: reported as near or above prior “frontier” closed models on several coding/agentic/security tasks, but:
    • It wins some benchmarks and loses others versus GPT-5.6 Sol, Opus 5, Kimi K3, GLM 5.3.
    • Several people caution about “benchmaxxing” vs real-world usefulness.
  • Early hands-on use: strong for coding, security analysis, and automated bug triage; some rate it slightly below top-tier closed models in overall “feel” but far above most open-weight models.

Hardware & local deployment

  • Full weights ~510–550GB plus KV cache; practical fast local use seen as requiring ~384GB+ RAM or multiple high-end GPUs.
  • 256GB systems can only run it with heavy quantization and/or SSD offload, which significantly slows prefill.
  • Sparsity and Engram offloading should make SSD streaming plausible for slow, unattended workloads.

Pricing & efficiency

  • Official pricing emphasizes very cheap cache hits (as low as $0.003–0.006 per 1M tokens off-peak), making long-horizon agentic workloads much cheaper than comparable closed models.
  • Relative to earlier V4 Flash and V4 Pro, pricing is somewhat confusing across providers; direct DeepSeek pricing for input tokens dropped vs recent V4 Flash peaks, output roughly similar, cache much cheaper.
  • Several users report multi-billion-token runs for a few tens of dollars.

Safety, openness & policy

  • Strong enthusiasm for open weights and detailed technical reporting, contrasted with US labs’ heavier focus on safety/“model welfare” documentation and closed weights.
  • Debate over safety vs “brand safety” and regulatory capture; some want fewer guardrails for security/pentesting, others argue alignment is under‑appreciated.
  • DeepSeek via some routes trains on user prompts, which some see as acceptable for coding, others avoid for sensitive or proprietary work.

Developer experience & tooling

  • DeepSeek Harness is praised as a powerful agentic coding environment; some prefer it to third-party harnesses.
  • V4.1 Flash will replace V4 Pro in routing; some already promote it as their primary “workhorse” for software factories.
  • Minor annoyances: Chinese language replies in the official UI for some users; occasional refusal behavior creeping in on “piracy-like” requests.