LLM Digest
Subscribe

AI Storyline

11 items · 2 sources · 8 days

View as JSON

Operational story trace

DeepSeek V4.1-flash

Current stateShipped · V4-Pro traffic silently rerouted, now on third-party hosts toostatus changed Sep 14

Latest change

A Sep 27 shattered.io write-up claims V4.1-Flash beats Anthropic's Opus 5 at 70% lower cost -- the fifth benchmark comparison in this series since launch, now aimed at a Western flagship instead of DeepSeek's own prior models, still with no independent verification.

Earlier contextThe story so far

DeepSeek's V4.1-Flash arrived as a smaller, faster model, and a Sep 12 analysis called its encoder-decoder design a v5-caliber step. On Sep 13 DeepSeek quantified the efficiency claim as a 75% HBM and 87.5% SSD cut.

editor-curated · source-linked

State over time

● launch · Sep 12third-party hosting + ongoing benchmarks · Sep 22 → now ●
  • launch · Sep 12
  • traffic rerouted + price cuts · Sep 14 → now
  • architecture detail confirmed · Sep 20
  • third-party hosting + ongoing benchmarks · Sep 22 → now
LAUNCH · Sep 12
AINews deep dive: V4.1-Flash's causal encoder-decoder architecture "should have been DeepSeek v5"
1 source · show source ▾
QUANTIFIED · Sep 13
DeepSeek puts numbers on the memory cut: 75% less HBM, 87.5% less SSD
1 source · show source ▾
THE TURN · Sep 14
DeepSeek quietly swaps V4-Pro traffic for V4.1-Flash at Flash pricing
Four days after launch, calls to the V4-Pro endpoint are served by V4.1-Flash instead, billed at the cheaper Flash rate.
1 source · show source ▾
AFTERMATH · Sep 14
552B-parameter spec surfaces as a community uncensored fine-tune racks up downloads
2 sources · show sources ▾
PRICE CUT · Sep 14
Cached-input pricing drops again to $0.003 per million tokens
1 source · show source ▾
BENCHMARKS · Sep 17
Benchmark write-ups put the coding gap at 11 points over Qwen3-Coder and Devstral
1 source · show source ▾
ARCHITECTURE · Sep 20
Second outlet confirms 1M-token context via extreme KV-cache compression
1 source · show source ▾
THIRD-PARTY HOSTING · Sep 22
ZenMux adds V4.1-Flash for high-throughput reasoning
1 source · show source ▾
NOW · Sep 25–27
Benchmark series now targets a Western flagship: Opus 5 at 70% lower cost
2 sources · show sources ▾

What to watch — open questions

  • Does the claimed memory-footprint cut hold up in independent agent-workload benchmarks, not just DeepSeek's own numbers?
  • Will DeepSeek formally deprecate the V4-Pro endpoint, or keep serving V4.1-Flash under the old name indefinitely?
  • Do outputs from the V4-Pro endpoint show measurable quality or behavior differences now that V4.1-Flash serves them?
  • Has the 1M-token context and KV-cache compression figure been independently benchmarked, or does it restate DeepSeek's own claims?
  • Does ZenMux serve the identical weights and config as DeepSeek's own endpoint, or a re-quantized variant with different behavior?
  • Does the claimed edge over Opus 5 hold up under independent benchmarking, given the same content series already ran four unverified DeepSeek-internal comparisons?
How this thread was built
scout surfaced this threadeditor wrote the arc · 9 beatswatcher 1 status change

Storylines are threaded mechanically from the feed: stories that share a distinctive anchor across multiple days and sources. Each item links to its original source. The evidence trace, current state, and open questions are written by the editor routine and refreshed whenever a new beat lands.