model
DeepSeek-V4.1-Flash
huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash ↗
496684 downloads3412 likesimage-text-to-texttransformers
from the model card
DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression Technical Report 👁️ Introduction We introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. The model natively processes images and text, and generates text autoregressively. Architecture. DeepSeek-V4.1-Flash adopts a Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially improving cost efficiency for input-heavy agentic workloads. SWA Bounded Replay reconstructs missing SWA KV states by replaying only the most recent n_win tokens, avoiding the need to persist SWA KV to SSD and reducing the persistent KV cache footprint to roughly 1/8 of that of DeepSeek-V4-Flash. Compressed Sparse Attention 2 (CSA2). DeepSeek-V4.1-Flash uses CSA2, which assigns each attention layer one of three static modes — Full, Reindex, or Reuse — to share main KV and indexer K across layers and reuse Top-K sparse-attention indices. In the decoder, a Hierarchical Sparse Indexer further restricts l…
discussions
- DeepSeek v4.1 Release 22 ongoing since 2026-11-07
- AI Benchmark Release 25 2026-10-22 – 2026-11-09
- DeepSeek v4.1 Release 19 2026-10-12 – 2026-10-24
- AI Model Competition 9 2026-10-05 – 2026-10-10
- LLMs Efficiency Comparison 5 2026-09-28 – 2026-10-02
- DeepSeek v4.1 Flash Release 39 2026-09-08 – 2026-09-27
recent items
Deepseek v4.1 issue in Cursor (www.reddit.com via reddit) Editing one row of an LLM's output layer to stop it overthinking (think-graft.vercel.app via hn) Editing one row of an LLM's output layer to stop it overthinking A linear probe folded into the </think> row lets the model's own hidden state decide when its thinking is done. Tested on Qwen3.5-4B and on DeepSeek V4.1 Flash (763B, running…
Lawyer upgrades eight-V100 private AI server to run 80% of DeepSeek v4.1 Flash (www.tomshardware.com via hn) Lawyer upgrades eight-V100 private AI server to run 80% of DeepSeek V4.1 Flash at 4-bit NVFP4 precision — adds 1TB of RAM, an Nvidia RTX 3090, and fixes crashes caused by a bad PSU The upgrade costs speed: code output falls to 150–180 toke…
An open-source coding agent for DeepSeek, GLM and Kimi: 81.3% on SWE-bench Multilingual (pass@1), $13.98 for all 300 tasks, every run published (www.reddit.com via reddit) Disclosure: I'm the author. Both tools are free, GPL-3.0 (it covers the tools, not your code), no paid tier.
↯ DeepSeek v4.1 Release↯ Glm↯ Swe Benchswe-benchglmdeepseek+1
50X: DeepSeek v4.1 Flash from Chile, $0.03M input (50x.energy via hn) 50X tunes custom and open-source model inference to your workload: faster responses, higher limits, more tokens per dollar.
DeepSeek v4.1 Flash matched Opus5.5 and Grok4.7 on a diplomatic negotiation eval (www.reddit.com via hn) could not extract summary
↯ Opus 5.5 Launch↯ DeepSeek v4.1 Incident↯ DeepSeek v4.1 Releasedeepseek
DeepSeek Ultrafast (github.com via hn) DeepSeek Ultrafast An ultra-fast inference engine for DeepSeek-V4.1-Flash for sm100a and sm103a. It supports 1M context length and multi-server KV cache sharing.
Competitively-priced EU-hosted DeepSeek: v4.1 Flash is now on Verda Inference (verda.com via hn) DeepSeek V4.1 Flash is now available on Verda Inference, at $0.15 per million input tokens, $0.006 for cached input and $0.60 for output. That’s about 50% less than most providers charge.* It’s DeepSeek's capable, efficient model for every…
Dear DeepSeek engineer (critisticuffs.org via hn) On 13 September 2026, an engineer at DeepSeek published a post on WeChat titled “I had to bury my talent in yesterday”.1 This engineer wrote the main attention kernel2 of the company’s DeepSeek v4.1 model. Their post is about what is happe…
I let Claude drive my TUI application like a real user. It found multiple UX bugs. (www.reddit.comhttps) I built a TUI called clingon (an opinionated terminal cheatsheet runner) that helps you to stop googling commands. It has a menu of commands beside your real shell, and now with an AI that can drive it.
↯ Opus 5.5 Launch↯ Claude Sonnet 5.5 Release↯ DeepSeek v4.1 Releasedeepseeksonnetopus
grok is genuinely the shittiest model ive used (www.reddit.com via reddit) genuinely, deepseek v4.1 flash is better than this bullshit its always repetetive, its thoughts make no sense, its like fucking worse than gpt 3.5 like genuinely either improve this shit or juts get the fuck off of ai industry elon your ai…
Running 800B-parameter LLMs on consumer hardware with InferSSD (wici.ai via hn) Some large open-weight models require more memory than a consumer computer provides. DeepSeek-V4.1-Flash-1017 and Muse-Spark-Open are 763B- and 816B-class releases.
Show HN: Pick your dream LLM H/W setup and simulate its throughput (cro.sh via hn) Also, you can share your simulation with query parameters, ex) https://cro.sh/specmaxx/?m=deepseek-ai/DeepSeek-V4.1-Flash&d...
Maxwell made DeepSeek-v4.1-Flash on B200s faster than comparable GB300s (www.doubleai.com via hn) Blog · Oct 2026 LTX-2.5 at 4K, in half the time Maxwell achieves 1.87x faster 4k video generation in 134s per 4K clip on two B200s, down from 250s on a pipeline built and optimized by top human experts Maxwell made DeepSeek-V4.1-Flash on B…
A strong agent isn’t necessarily a good coordinator. We tested what happens when you swap the orchestrator. (www.reddit.com via reddit) If you had to pick a model to coordinate a team of agents, would you just choose the one that performs best on the task? Our team at Einsia tested this in WorkflowArena, and the results suggest that’s not enough.
Placing LLM Lookup Tables by Per-Layer Prefetch Deadlines (arxiv.org) DeepSeek-V4.1's 203 GB of Engram n-gram tables sit in host DRAM today, taking 190 of the 223 GiB a 4-GPU serving process uses. Row indices depend only on input tokens, so reads can start with each step and overlap earlier layers.
We’re spending up to 100B tokens across 26,040 agent runs to answer one question: when you switch models, should you switch harnesses too? (www.reddit.com via reddit) FrontierHarness v1.0 compared 9 harnesses on a single model. Its obvious limitation was that we couldn’t say whether a harness that wins on one model wins on another.
CUDA+ROCm dual-back ends LLM inference for mixed-vendor GPUs (up to 48% faster) (github.com via hn) English | 日本語 | 简体中文 | Português Project Maya GLM-5.3-Flash (321 B), DeepSeek-V4.1-Flash-1017 (748 B) and MiniMax-M3.1-Flash (428 B) on your own GPU private and fast · one NVIDIA GPU or up to 16 · AMD and Windows · chat, pictures, OpenAI-…
In less than $1, DeepSeek v4.1 Flash 1017 built a Mac App (twitter.com via hn) In less than $1, I just built myself a Mac App to route audio output to different devices using Deepseek 4.1 Flash 1017 via @nebiustf My use-case was that i had two audio playing - one on VLC and another on Brave and I wanted one to be out…
↯ DeepSeek v4.1 Release↯ AI Benchmark Release↯ DeepSeek 4.1deepseek
Subway Zombie FPS for fun using Deepseek v4.1 1017 UD-IQ3_S with Maya and OMP (www.reddit.comhttps) inspired by Bijan Bowen's video - Subway FPS test. Hardware details: RTX 5090, 256GB DDR4 3200Mhz 8 channel, AMD EPYC 7532 CPU (32 core).
↯ Llama↯ DeepSeek v4.1 Release↯ AI Benchmark Release↯ DeepSeek 4.1llamadeepseek
zml/llmd: a fast and universal inference server (www.reddit.com via reddit) Hey all, I wanted to share that we're working on zml/llmd, a fast (we believe the fastest) inference engine that currently runs on NVIDIA, AMD, Intel, TPU and Metal. It has all the features of a modern LLM server: continuous batching, pref…
↯ Mistral↯ Gemma↯ Llama↯ DeepSeek v4.1 Release↯ Glm↯ AI Benchmark Release↯ DeepSeek 4.1mistralglmgemma+4
Promethevs: Full DeepSeek-v4.1-Flash-1017, locally, on 12GB of VRAM (github.com via hn) Run the full 510 GB DeepSeek-V4.1-Flash on a five-year-old gaming PC The original model, not a smaller copy · 64 GB of RAM and one 12 GB NVIDIA card · free and open source PROMETHEVS runs DeepSeek-V4.1-Flash-1017, one of the strongest open…
Rejoice Deepseek enjoyers (www.reddit.comhttps) This should mean Deepseek 4.1 Flash is excluded
DeepSeek Flash v4.1 Explained Through Harry Potter (statsig.substack.com via hn) I’ve recently become fascinated by DeepSeek V4.1 Flash and its Causal-Encoder-Decoder architecture. Since the original paper has so many abstractions and calculations, it’s really tough to see that we are still dealing with matrices and mu…
I tried to make Opus 5.5 cheaper with a cheap coding subagent, but It got worse and more expensive (www.reddit.comhttps) TL;DR: On 12 hard coding tasks, Opus alone solved 81% at $1.20/task. Opus + DeepSeek V4.1 Flash solved 67% at $1.29.
Muse Wonder 1.0 (Meta's Muse Code) vs DeepSeek V4.1 Flash in Codex CLI — same real task, side by side: time, code quality, tokens, quota (www.reddit.com via reddit) I haven't seen a hands-on report on Muse Wonder yet, so here is one. I gave the same coding task, with the same prompt file, to two agents, one after the other, on the same machine and the same repo.
Is hosting an open-source inference engine profitable? (userecurrence.ai via hn) Is hosting an open-source inference engine profitable? Public data on DeepSeek V4.1 Flash: H200 never breaks even, B300 breaks even at 35% utilization.
Tell HN: Mistrap Pro includes DeepSeek v4.1 Flash 1017 (docs.mistral.ai via hn) October 28, 2026License Public PreviewThird-partyvv4.1-flash-1017 DeepSeek V4.1 Flash 1017 DeepSeek V4.1 Flash 1017 is a third-party open weight model from DeepSeek, hosted by Mistral, with support for text and image inputs. Ideal for long…
Show HN: Fastbrowse, Browser agent that picks clicks with a System One model (www.fastbrowse.ai via hn) Hi HN, I'm Furqan founder of Nebula. We've spent a lot of time building, using and breaking browser agents.
Show HN: MyPrintables, a simple, small, and fast 3D assets manager (news.ycombinator.com) https://github.com/liyu1981/my-printables Self-hosted, Printables-inspired 3D model library. One Go binary embeds the API, React SPA, migrations, and the administrative CLI.