DeepSeek V4.1 Flash rumored at 552B params as Engram memory beats matched MoE
teortaxesTex · x · 2026-11-08
Unverified speculation by teortaxesTex connecting DeepSeek and Qwen model leaks:
- "Qoduck" is guessed to be Qwen4-27B (35GB in Q8, 480 tps, extremely verbose like the old Qwen 35B 3AB);
- Qwen 3.8-Flash-next was an architecture preview for Qwen4;
- DeepSeek's original Engram paper tested a 27B-total/3.8B-activated MoE, but with a different naming convention;
- V4.1 Flash's specs are "552B total + 196B activated," closer to a "Flash-next" scale.
The attached image is Table 1 from the Engram paper: at 262B tokens with 3.8B activated params, Engram-27B (reallocating routed experts 72→55 into a 5.7B Engram memory) beats matched MoE-27B across MMLU, C-Eval, ARC and other benchmarks, with Engram-40B (18.5B memory) leading further.
Related event: Leaked Qwen4-27B Early Checkpoint Tops GameFeel Bench at Just 35GB(4 posts)→
More from Models
- Garbage in, garbage out: Opus makes bad videos from bad starting points — pramodk73 · 2026-11-08
- Counterintuitive: Opus 5.5 fails 2x more turns on max than on high, in every time bucket — tristanbob · 2026-11-08
- Thinking Machines serves Inkling 2 Small free on OpenCode for a limited time to gather feedback — clarejtbirch · 2026-11-08
- MaxxBench strips away benchmaxxing: Claude Opus 5.5 still tops at 74.2% — craterIII · 2026-11-08
- Claude wrote a wrong Pokémon landmark note and kept trusting it — Anthropic tested a memory repair pass — shoumikgoswami · 2026-11-08
- Sonnet 4.6 rates models on self-designed scales; author says 'self-report' caveats are performative — Sauers_ · 2026-11-08