DeepSeek Releases Open-Source V4.1 Flash, Beating Its Own Flagship at a Fraction of the Cost
DeepSeek officially released and open-sourced V4.1 Flash on September 10 across web, app, and API, with weights on Hugging Face under the MIT license (fp8/8-bit builds, transformers support) plus an accompanying paper. The company claims it beats its own flagship V4 Pro on performance, cost, and speed, making it one of the most-watched open-source releases of the season.
Confirmed
- Architecture: A 552B-parameter asymmetric MoE with a new Causal Encoder-Decoder design—8B parameters active on input, 16B on output—and up to 1M-token context. Cross-layer shared compressed KV cache saves 7/8 of KV cache cost (@YuchenjUW notes a further 4x per-token KV reduction), plus a two-stage sparse indexer.
- Multimodal & ecosystem: First model in the new architecture family with native vision (image-text-to-text), and the smallest member so far, paving the way for larger siblings. vLLM and SGLang offer day-zero support verified on NVIDIA/AMD GPUs; DeepSeek is soliciting 2000-GPU-plus-storage cluster-scale deployment partnerships.
- Pricing & benchmarks: New pricing with 50% peak/off-peak gap; community reports $0.14-0.15/M input and $0.60/M output. @kimmonismus notes output price is 6% of GPT-5.6 Sol's; @YuchenjUW puts total cost near 3%. Per Artificial Analysis, it scores 40 on the Intelligence Index—beating DeepSeek's 1.6T Pro flagship—at roughly 4x lower token price. Community posts cite DeepSWE v1.1 at 74.2%, CyberGym 88.1 vs 84.5, and API cost around 2.5% of Opus 5.
- V4 Pro retirement: Until V4.1 Pro launches, all V4 Pro requests are auto-routed to V4.1 Flash at lower prices; per @markk, routing starts September 14. V4-Flash and V4-Flash-Vision-E are also being retired. This comes only 6 weeks after July's V4-Flash.
- Reception: Hugging Face co-founder Thom Wolf says it tops the open-source leaderboard at rock-bottom prices and shared a forward-pass visualization video; antirez is more skeptical, seeing only a marginal gain over 4.0 Flash despite doubled backbone parameters.
Unconfirmed
- Some reposts (@solyarisoftware) cite 769B total parameters, contradicting the official 552B figure. @ns123abc's claims of 1/86 competitor cost and 420-507 tok/s output speed stem from unverified leaks.
Why it matters
- Matching or beating a flagship at a fraction of the cost, fully MIT-licensed, could further compress industry inference prices and pressure closed vendors—though the generational jump over its predecessor, as antirez notes, remains debated.
- V4 Pro users should watch the September 14 auto-routing cutover and evaluate real-world behavior beforehand.
- 1M-token context, native vision, and 7/8 KV cache savings directly affect long-document and agentic workflows, and the model sets the stage for larger siblings in the same family.
2026-09-09 ~ 2026-09-11 · 57 related posts
- 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)
Primary sources
- DeepSeek routes V4 Pro requests to faster, cheaper V4.1 Flash, hints V4.1 Pro is coming — zephyr_z9 · 2026-09-09
- DeepSeek releases V4.1 Flash: 22% cheaper than V4 and better performance — deliprao · 2026-09-09
- DeepSeek ships V4.1 Flash: 552B MoE on new architecture, MIT-licensed — deepseek-ai · 2026-09-10
- DeepSeek releases V4.1 Flash: 552B MoE beats V4 Pro, cuts prices, retires flagship — DeepSeek · 2026-09-10
- [source] DeepSeek Launches V4.1-Flash, Smallest Model in New Architecture Family with Native Vision — deepseek_ai · 2026-09-10
- DeepSeek Details V4.1-Flash: 552B MoE with Asymmetric 8B/16B Active Params, Beats Its Flagship — deepseek_ai · 2026-09-10
- [source] DeepSeek-V4.1-Flash Hits the API: 1/8 the SSD Cache, V4-Pro Being Phased Out — deepseek_ai · 2026-09-10
- [source] DeepSeek Open-Sources V4.1-Flash Under MIT License, Cuts Off-Peak API Rates to 50% — deepseek_ai · 2026-09-10
- DeepSeek V4.1 Flash drops: 552B MoE activating just 8B, beating GPT-5.6 Sol at $0.14/M — ChrisGPT · 2026-09-10
- DeepSeek to retire v4 Pro, reroute all requests to new v4.1 Flash — gaganghotra_ · 2026-09-10
- DeepSeek releases V4.1 Flash: 552B MoE, 1M context, SGLang day-0 support — BanghuaZ · 2026-09-10
- Inside V4.1 Flash: Per-Token KV Cache Down to 890 Bytes, Peak-Valley Pricing — 赛博禅心 · 2026-09-10
- DeepSeek-V4.1-Flash appears on Hugging Face: MIT-licensed multimodal model with FP8 weights — AIFlow_ML · 2026-09-10
- DeepSeek V4.1 Flash weights go live on Hugging Face under MIT license — solyarisoftware · 2026-09-10
- DeepSeek V4.1 Flash now live on App, Web and API with open weights — gaganghotra_ · 2026-09-10
- Perplexity CEO reacts to DeepSeek V4.1 Flash launch — keviv9 · 2026-09-10
- DeepSeek V4.1 Flash debuts: 769B MoE with native vision, 1M context, day-0 vLLM support — solyarisoftware · 2026-09-10
- DeepSeek previews new architecture: 552B MoE with just 8B active input parameters — zephyr_z9 · 2026-09-10
- DeepSeek V4.1 Flash: 552B MoE with just 8-16B active params, nearing GPT-5.6 at a fraction of cost — brainlatch42 · 2026-09-10
- DeepSeek V4.1-Flash: New Architecture, Beats GPT-5.6 Sol on DeepSWE at 94% Lower Output Price — kimmonismus · 2026-09-10
- Leaked: DeepSeek V4.1 Flash open weights, claims 86x cheaper and 420+ tok/s — ns123abc · 2026-09-10
- SGLang and Miles ship day-0 support for DeepSeek-V4.1 Flash and detail its new architecture — ying11231 · 2026-09-10
- DeepSeek launches V4.1-Flash with native vision; antirez says 2x params beat 4.0 Flash only modestly — solyarisoftware · 2026-09-10
- DeepSeek V4.1 Flash tops open-source leaderboard with extreme efficiency — Thom_Wolf · 2026-09-10
- DeepSeek-V4.1-Flash lands on Hugging Face: 552B params, 1M-token context — victormustar · 2026-09-10
- DeepSeek Launches V4.1-Flash; Dettmers Says It's as Big as R1 — PMinervini · 2026-09-10
- DeepSeek Launches V4.1-Flash at ~2.5% of Opus 5's Cost, Beating Flagships on DeepSWE — ccerrato147 · 2026-09-10
- DeepSeek Details V4.1-Flash: 552B MoE with Just 8B Active Params on Input — ccerrato147 · 2026-09-10
- DeepSeek-V4.1-Flash released: open weights trade blows with GPT-5.6 and Claude Opus 5 — AccBalanced · 2026-09-10
- DeepSeek V4.1 Flash beats GPT-5.6 Sol on coding while costing ~97% less — Yuchenj_UW · 2026-09-11
20 near-duplicate retellings: 机器之心 · eliebakouch · reach_vb · teortaxesTex · basedjensen · AI寒武纪 · intellectronica · Scobleizer · Scobleizer · APPSO · ricklamers · zephyr_z9 · 量子位 · syhw · NVIDIAAI · joecole · dair_ai · solyarisoftware · mark_k · alexcovo_eth