DeepSeek v4.1-Flash buys #12 rank by moving a 196B Engram off HBM

TL;DR

• DeepSeek v4.1-Flash shrinks KV cache to 890 bytes per token, a 437× drop from V1.
• A 196B-parameter Engram table sits in DDR5/NVMe, keeping HBM off the critical path.
• Independent benchmarks place it at #12 overall (51.8) at $0.30/M input tokens.
• Users report linguistic drift and reasoning-token bloat that erode the price advantage.
• Anthropic alleges ~12M Claude queries from ~24,000 DeepSeek-linked fraudulent accounts in July.

Today’s research slate is a single feature, and its interest is architectural rather than competitive. DeepSeek v4.1-Flash does not chase the frontier — it lands at rank #12 on independent benchmarks — but it gets there by rewriting where the model’s memory lives. A 437× KV cache reduction to 890 bytes per token, plus a 196B-parameter Engram N-gram table pushed into DDR5 and NVMe, mean the 763B/8B–16B-active math runs on rigs that are not hyperscale.

The catch is that the same relocation surfaces in production as linguistic drift and reasoning-token bloat, chewing back into the $0.30/M input price advantage. And Anthropic’s claim of ~12M Claude queries from ~24,000 DeepSeek-linked fraudulent accounts hangs over the release as a reminder that model economics and account-abuse economics are increasingly the same story.

DeepSeek v4.1-Flash cuts KV cache to 890 bytes per token

Source: latent-space · published 2026-09-12

TL;DR

• DeepSeek v4.1-Flash shrinks its KV cache to 890 bytes per token, a 437× reduction from V1.
• A 196B-parameter Engram N-gram table lives in DDR5/NVMe, not HBM — making the 763B/8B–16B-active math run on non-hyperscale rigs.
• Independent benchmarks put it at #12 overall (51.8) at $0.30/M input — a price-tier win, not a frontier crown.
• Practitioners flag linguistic drift and reasoning-token bloat that erode the price advantage in production.
• Anthropic alleges ~12M Claude queries from ~24,000 DeepSeek-linked fraudulent accounts over two weeks in July.

The 437× cache shrink is the real story

The version bump undersells what shipped. An independent teardown by local-ai-zone documents three stacked tricks behind v4.1-Flash’s 890-byte-per-token KV footprint: Compressed Sparse Attention 2 assigns each layer a static Full / Reindex / Reuse mode so KV data and sparse indices are shared across layers; FP4 (E2M1) caching applies per-16-channel scaling; and a sliding-window “bounded replay” pass reconstructs SWA state instead of persisting it ¹. Together that’s a 437× cache shrink from DeepSeek V1 ¹ — the kind of number that changes what fits on a single node, not just what fits in a slide.

The other half of the “763B total, 8B–16B active” claim (the P8B-D16B in the launch title) comes from Engram, the conditional-memory module. Introl’s write-up notes that because Engram is accessed via deterministic hashing rather than learned routing, its 196B parameters don’t need to live in HBM — they can be memory-mapped to host DDR5 or local NVMe ². That’s the detail that makes the parameter budget honest: only the dense trunk and active experts actually touch GPU memory.

flowchart LR
    Q[Token / query] --> R{Deterministic hash}
    R --> H[HBM: dense trunk + active MoE experts ~8-16B]
    R --> D[DDR5 / NVMe: Engram 196B N-gram table]
    H --> A[Attention w/ FP4 KV cache · 890B/token]
    D --> A
    A --> O[Output]

Price-tier win, not a frontier crown

LLM-Stats’ aggregate has v4.1-Flash at #12 overall with a composite score of 51.8, ahead of Qwen 3.7 Max (45.5) but well shy of the proprietary frontier — at $0.30 per million input tokens, roughly 5.7× cheaper than Qwen, with an additional 50% off-peak discount ³. That explains DeepSeek retiring the Pro tier by routing traffic to Flash, but the “Return of the Whale” framing needs an asterisk. Coursiv’s industry survey argues the market has bifurcated: Mamba/SSM hybrids like AI21’s Jamba, NVIDIA’s Nemotron-Flash and Microsoft’s SambaY own streaming and edge, while the MoE-Transformer CED stack only wins clearly on genuinely asymmetric agentic reasoning ⁴.

Live failure modes practitioners keep hitting

Developer reviews surface two ergonomics regressions vendor benchmarks don’t. The first is linguistic drift — reasoning chains that flip between English and Chinese mid-turn ⁵. The second is more expensive: an “annoyingly overthinking” mode that burns reasoning tokens on trivial prompts, which matters a lot if you’re not on the off-peak discount ⁵. Both are the kind of thing that make a 5.7× price advantage evaporate in production.

The distillation cloud

The launch also lands mid-controversy. Anthropic’s September 2026 threat-intel report alleges roughly 12 million Claude exchanges from ~24,000 fraudulent accounts routed through proxy “transfer stations” tied to DeepSeek over a two-week July window ⁶. DeepSeek credits v4.1’s jump to CED plus a 45T-token corpus heavy on synthesized agent trajectories, but the provenance question now sits on top of every Claude- or GPT-relative comparison. The architecture is genuinely novel and unusually well-documented ¹²; whether the training data is equally clean is a separate — and unresolved — question.

Footnotes

1. local-ai-zone deep dive (github.io) — https://local-ai-zone.github.io/blog/deepseek-v4-1-flash-deep-dive.html

   Compressed Sparse Attention 2 (CSA2)… Each layer is assigned a static mode — Full, Reindex, or Reuse — allowing the model to share KV data and reuse sparse-attention indices across layers… FP4 KV Caching uses an E2M1 format with per-16-channel scaling, shrinking the global KV cache to just 890 bytes per token, a 437x reduction from DeepSeek V1.

   ↩ ↩² ↩³

2. Introl blog on Engram conditional memory — https://introl.com/blog/deepseek-engram-conditional-memory-architecture-january-2026

   Because the Engram module is sparsely accessed via deterministic hashing, the 196B parameters do not need to reside on GPU HBM; they can be memory-mapped and offloaded to host DDR5 RAM or local NVMe storage.

   ↩ ↩²

3. LLM-Stats comparison page — https://llm-stats.com/models/compare/deepseek-v4.1-flash-vs-qwen3.7-max

   v4.1 Flash ranks #12 overall with a composite score of 51.8, outperforming Qwen 3.7 Max (45.5)… priced at $0.30 per million input tokens, roughly 5.7x cheaper than Qwen 3.7 Max, with an additional 50% off-peak discount.

   ↩

4. Coursiv industry-trend piece — https://coursiv.io/blog/deepseek-v4-1-flash

   The market has bifurcated: Mamba/SSM hybrids from AI21 (Jamba), NVIDIA (Nemotron-Flash) and Microsoft (SambaY) offer superior throughput for streaming and edge, while DeepSeek’s optimized MoE-Transformer CED stack remains more competitive for complex, multi-step agentic reasoning.

   ↩

5. YouTube developer review — https://www.youtube.com/watch?v=Weom9fQnnJ0

   Linguistic drift where the model’s thinking chain unpredictably switches between English and Chinese… a vocal subset labeled the model ‘annoyingly overthinking,’ burning excessive reasoning tokens on simple tasks and driving up cost for users not on off-peak pricing.

   ↩ ↩²

6. BigGo finance / Anthropic threat report coverage — https://finance.biggo.com/news/2f97922b-45b6-4f7c-9d6d-adcf66696602

   Anthropic’s September 2026 threat intelligence report detailed over 12 million exchanges linked to DeepSeek in a two-week window in July 2026, using approximately 24,000 fraudulent accounts and proxy ‘transfer stations’ to query Claude for step-by-step reasoning.

   ↩