DeepSeek V4.1 Flash Leaks: 552B Asymmetric MoE Claimed to Rival GPT-5.6 Sol

Starting September 10, multiple posters leaked claimed official benchmarks and technical details of DeepSeek V4.1 and V4.1 Flash, with Sentdex reporting the next day that weights had been released. If true, V4.1 would challenge GPT-5.6 Sol on agentic/coding benchmarks with a 552B asymmetric-activation architecture, dramatically cut inference memory and cost via extreme KV-cache compression, and introduce a rare 1-100 continuous reasoning-effort control. Nothing is officially confirmed, and iScienceLuvr among others suspects benchmark gaming.

Confirmed

• All information is unofficial leaks, from teortaxesTex, eliebakouch, ChrisGPT, nrehiew, iScienceLuvr, AdinaYakup, zainhas, Sentdex, @sheriyuo (via AndrewZhao) and @MiaAIlab (via airesearch12), with figures cross-corroborated; SemiAnalysis also posted "congratulations" (via bodonoghue85).
• Total parameters 552B with a new Causal-Encoder-Decoder architecture: 8B input activation / 16B output activation; nrehiew notes the same model activates different parameter counts across prefill and decode stages, an unusual design. Sources stress costs well below same-size models; sheriyuo claims it is smarter than DeepSeek's own flagship; MiaAIlab says capability is on par with or better than GPT-5.6 Sol, Claude Opus 5, GLM 5.3 and Kimi K3; Sentdex's cited post says weights are out and it beats Opus-5, GPT-5.6-Sol, Kimi-K3 and GLM-5.3 in key areas.
• Leaked benchmarks: TerminalBench 4.0 at 31.2, CyberGym 88.1 (vs GPT-5.6 Sol's 84.5), HLE 63.9; ChrisGPT claims leads across agentic/coding benchmarks; nrehiew reports V4.1 Flash tops DeepSWE at 74.2, above Opus 5 and Gemini 3.8 Flash.
• Per eliebakouch, V4.1 Flash is a natively multimodal model trained on 45T tokens, with an architecture praised as the most novel in years; AdinaYakup notes native vision is merged into one endpoint.
• zainhas highlights reasoningeffort continuously adjustable from 1 to 100, rather than discrete low/medium/high tiers, plus a rare API compatibility mapping — a design they say they have never seen before.

Architecture and engineering highlights

• Per teortaxesTex, KV cache compressed to 890 bytes/token (under 1GB per million tokens); HBM demand down another 3.9x and SSD down 8x versus V4-Flash. AdinaYakup cites KV cache at 1/437 of the first generation, HBM at 1/4 and SSD at 1/8 of the previous generation.
• Scobleizer's relay adds pricing: $0.15 input / $0.60 output per million tokens; AdinaYakup calls it one-fortieth of Opus's price.
• Also per teortaxesTex (via zephyrz9), V4.1 removes V4's stopgaps: drops HCA, generalizes CSA into a more general primitive, removes sparse-attention warmup, and adopts a simpler single-pass mHC; a new sparse attention design is also mentioned (via donglixp).
• teortaxesTex also jokingly relayed a report signed "fruit fly connectome" claiming V4.1-Flash inference optimized on 2× NVIDIA GB300 DGX Stations with 56k tok/s prefill and 201.7 tok/s c1 decode; its credibility is dubious by nature.

Unconfirmed

• All benchmarks, specs, pricing, throughput figures, architecture details and the weights-release claim lack official sources; iScienceLuvr explicitly doubts the "GPT-5.6 Sol level" scores, suspecting benchmark gaming, with "benchmarkmaxxed" skepticism appearing in related threads.

Why it matters

• If true, V4.1 goes head-to-head with GPT-5.6 Sol on agentic/coding while asymmetric activation and extreme KV compression slash deployment memory barriers; the 1-100 continuous reasoning effort would offer a new interaction paradigm for API-side inference cost and depth control, with significant cost implications in an open-weights ecosystem.

2026-09-10 ~ 2026-09-11 · 21 related posts

Full story(18 episodes)→

• Episode 1: DeepSeek V4.1 Flash Opens Limited Internal Beta with New Architecture and Native Multimodality(2026-09-08, 18 posts)
• Episode 2: DeepSeek V4.1 Flash Tested: Blazing 350 Tokens/s but Still Experimental(2026-09-08, 2 posts)
• Episode 3: DeepSeek cuts V4-Flash API prices with new peak/off-peak billing from Sept 10(2026-09-08, 6 posts)
• Episode 4: DeepSeek V4.1 Flash Nears Frontier Models at a Fraction of the Cost, Independent Tests Show(2026-09-09, 15 posts)
• Episode 5: DeepSeek Releases Open-Source V4.1 Flash, Beating Its Own Flagship at a Fraction of the Cost(2026-09-09, 57 posts)
• Episode 6: DeepSeek V4.1 Flash Leaks: 552B Asymmetric MoE Claimed to Rival GPT-5.6 Sol(2026-09-10, 21 posts)
• Episode 7: DeepSeek V4.1 Flash Benchmark Results Leak, Sparking Community Discussion(2026-09-10, 2 posts)
• Episode 8: DeepSeek Shrinks Per-Token KV Cache 54x in Nine Months to Sub-KB Levels(2026-09-10, 6 posts)
• Episode 9: Leaked DeepSeek-V4.1-Flash Touted as New Open-Source King(2026-09-10, 3 posts)
• Episode 10: Bug Hunt Bench Ranks Frontier Coding Models on 105 Real Bugs(2026-09-10, 13 posts)
• Episode 11: DeepSeek V4.1 Flash Confirmed to Center on YOCO Architecture(2026-09-10, 7 posts)
• Episode 12: Leaked DeepSeek V4.1 Flash Tech Report Reveals KV Cache-Obsessed Flagship(2026-09-11, 44 posts)
• Episode 13: DeepSeek V4.1 Flash tops Vals open-source index at $0.30 per run(2026-09-11, 2 posts)
• Episode 14: DeepSeek launches V4.1-Flash: 552B backbone, 1M-token context(2026-09-11, 5 posts)
• Episode 15: DeepSeek V4.1 Flash scores 40 in AA index despite rising hallucinations(2026-09-11, 2 posts)
• Episode 16: Inside KV-Cache Sharing: How DeepSeek CED and GLM 5.2 Differ(2026-09-11, 4 posts)
• Episode 17: DeepSeek 4.1 Flash Shows Big Gains in Kernel Engineering and Cost(2026-09-11, 2 posts)
• Episode 18: SGLang delivers Day 0 support for DeepSeek V4.1 Flash, hitting 873 tok/s(2026-09-12, 2 posts)

• [source] Leaked DeepSeek V4.1 benchmarks show 552B new-architecture model hitting 63.9 HLE with tools — teortaxesTex · 2026-09-10
• DeepSeek V4.1 Leak: 552B New Architecture, HBM Needs Cut 3.9x, Huge Benchmark Gains — teortaxesTex · 2026-09-10
• [source] DeepSeek slashes KV cache to 890 bytes/token, hinting at million-agent swarms — teortaxesTex · 2026-09-10
• DeepSeek V4.1 Flash rumored: 552B MoE with asymmetric CED architecture, beats flagship — _AndrewZhao · 2026-09-10
• DeepSeek V4.1 Flash benchmarks leak: beats GPT-5.6 Sol on agentic/coding tasks at $0.14/M input — ChrisGPT · 2026-09-10
• DeepSeek V4.1 Flash leak: 552B MoE with 8/16B active, native vision, praised as most novel arch in years — eliebakouch · 2026-09-10
• Leak: DeepSeek V4.1 reportedly drops V4's HCA, generalizes CSA, removes sparse-attention warmup — zephyr_z9 · 2026-09-10
• DeepSeek V4.1 Flash hits 74.2 on DeepSWE, beating Opus 5 and Gemini 3.8 Flash; asymmetric MoE revealed — nrehiew_ · 2026-09-10
• DeepSeek-V4.1-Flash reportedly offers continuous reasoning effort from 1 to 100 — zainhas · 2026-09-10
• DeepSeek V4.1-Flash leak: 552B params claiming GPT-5.6-level scores via CED architecture — iScienceLuvr · 2026-09-10
• Rumored DeepSeek V4.1 Flash details point to asymmetric-activation MoE with much lower cost — gaganghotra_ · 2026-09-10
• DeepSeek-V4.1-Flash offers continuously controllable reasoning effort from 1 to 100 — zainhas · 2026-09-10
• 552B params but only 8B active: DeepSeek V4.1 Flash's wild efficiency numbers — Scobleizer · 2026-09-10
• DeepSeek V4.1-Flash: KV cache 437x smaller, ~40x cheaper than Claude Opus 4.8 — AdinaYakup · 2026-09-10
• Leak claims DeepSeek v4.1 Flash matches or beats GPT-5.6 Sol and Claude Opus 5 — airesearch12 · 2026-09-10
• DeepSeek-V4.1-Flash leaks: 552B backbone activating just 8B, 8x less KV cache — bodonoghue85 · 2026-09-10
• DeepSeek V4.1-Flash turns heads: GPT-5.6-level benchmarks at 552B params — teortaxesTex · 2026-09-10
• [source] DeepSeek V4.1 Flash weights out, outperforming Opus-5 and GPT-5.6-Sol in key areas — Sentdex · 2026-09-11
• Unverified DeepSeek-V4.1-Flash Benchmark Claims 56k tok/s Prefill on Dual GB300 — teortaxesTex · 2026-09-11

2 near-duplicate retellings: zainhas · donglixp