model
huggingface.co/deepseek-ai/DeepSeek-V4-Flash ↗
1991598 downloads1625 likestext-generationtransformers
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence Technical Report👁️ Introduction We present a preview version of DeepSeek-V4 series, including two strong Mixture-of-Experts (MoE) language models — DeepSeek-V4-Pro with 1.6T parameters (49B activated) and DeepSeek-V4-Flash with 284B parameters (13B activated) — both supporting a context length of one million tokens. DeepSeek-V4 series incorporate several key upgrades in architecture and optimization: Hybrid Attention Architecture: We design a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to dramatically improve long-context efficiency. In the 1M-token context setting, DeepSeek-V4-Pro requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2. Manifold-Constrained Hyper-Connections (mHC): We incorporate mHC to strengthen conventional residual connections, enhancing stability of signal propagation across layers while preserving model expressivity. Muon Optimizer: We employ the Muon optimizer for faster convergence and greater training stability. We pre-train both models on more than 32T diverse and high-quality tokens, followed by a comprehensive post-training pipeline. The post-training features a two-stage paradigm: independent cultivation of domain-specific experts (through SFT and RL with GRPO), followe…
Hey everyone, it's been a while since I updated anyone on the memory system I built for my AI assistant Friday. Well, I had Deepseek V4 Flash write up a system map for itself, and I figured that probably people here would be interested in…
↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4deepseek
OpenCode Senses can inspect screenshots, extract exact OCR, detect and locate objects, zoom into regions, compare two images, measure colors, crop and annotate images, and even reverse-search them. Everything runs locally, so it's private,…
↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4deepseek
could not extract summary
↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4deepseek
Rig: 3945WX (12C, 2 CCDs, no AVX-512) · 8×32GB DDR4-3200 · 4× 5060 Ti 16GB · PCIe 4.0. Agentic workload (Hermes Agent).
↯ Vllm↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4↯ DeepSeek 4vllmdeepseekagentic
As a DSH glazer, I'm not going to be biased and shield the actual model being used from criticism. I asked DSH to implemented a feature within DSH.
I'd be curious to try it locally since I use 0731 daily but still no news on the weights
The American AI industry built its business on one bet above all others: that the most capable models, kept behind an API and priced to match, would always stay a step ahead of anything a rival was willing to give away for free. This week,…
DeepSeek-V4-Pro is a 1.6-trillion-parameter Mixture-of-Experts (MoE) model released with both FP8 and FP4 weights. Models at this scale naturally benefit from accelerators such as NVIDIA Blackwell GPU...
At $10k, you could get - 6.2B tokens with Qwen 3.8 Max (Qwen Pro plan) - 5.7B tokens with DeepSeek V4 Pro OpenRouter - 100B tokens with DeepSeek V4 Flash OpenRouter As a firm believer of local inference, unless you need it for data soverei…
I’ve been experimenting with a custom inference/runtime research project called CRANE V2, mostly because I wanted to answer a stupid question: How far can you push absurdly large MoE models on an ordinary consumer Windows machine before ph…
Lv:1 飞 机 大 战 最高分：0 EN | 中 开始游戏 🖱️ / 👆 移动战机（自动开火） ←↑↓→ / WASD 键盘移动 B / 炸弹 释放全屏炸弹 P · M 暂停 · 静音 道具：蓝色小飞机=双发火力 · 红炸弹=全屏炸弹 · 蓝盾牌=护盾 · 深蓝僚机=伴飞战机（最多 3 架） 击杀 BOSS 掉落武器核心，可切换子弹形态（直射 / 散射 / 贯穿 / 侧翼）· 被击中时若持有炸弹会自动引爆保命 已 暂 停 继续游戏 返回主菜单 按 P 键或点击按钮继续 游…
My test is running DeepSeek v4 Flash 0731 at native size on DGX Spark machine (GB10, 128 GB unified memory). The model size is bigger than RAM, so weights will be loaded many times when running.
Hey all, So I currently use 2x DGX Sparks with Deepseek v4 Flash 0731 and it works fantastic at 1M context, with 1.8M kv, dspark, vllm tp 2, etc. All good.
I just finished building my (relatively) low rent local inference machine: * Epyc 7663 * 256GB ECC DDR4-3200 * 1x RTX 5090 32GB Yeah I realize it's weird to throw a 5090 and 256GB of anything together and call it low end, but relative to ~…
**Setup:** 2× DGX Spark (GB10, 121 GiB unified each), TP=2 over 2×200GbE RoCE, vLLM 0.25.2.dev0, DeepSeek-V4-Flash-0731 FP8, `max_model_len=450000`, prefix caching on. KV pool = **1,686,693 tokens**.
This is one more vibe slopped custom optimization for, in this case, my hardware (m2 ultra 60 cores, 192gb). It is just a fork from llama.cpp with a few changes, it achieves: - DeepSeek V4 Flash, no kv cache quant - 141GiB model, byte-iden…
Super new to this local LLM stuff. Just set up a 2x Asus Ascent GX10 cluster and have DeepSeek V4 Flash 0731 running on it.
The post describes some experiments I had while trying to desperately run deepseek-v4-flash-0731 4 bit+ quants on my machine which is supposed to support only q2 quants of the model, a or 2.xx bpw quants at best. Long story short , I wante…
So I wanted to check out some of the current models in a repeatable benchmark, so I thought I'd share the result with you all. Models in this test: Ling 3.0 Flash Q4_K_M, Ornith 1.5 35B Q8_0, Deepseek V4 Flash UD_Q2_K_XL, Nemotron 3.5 Ligh…
I was experimenting with deepseek harness when found that even if you don't use deepseek models, you can configure the web_search tool with their api key and every hit will cost you as if you called deepseek-v4-flash model. It's a bummer.
https://preview.redd.it/84zi5nsdawkh1.png?width=2368&format=png&auto=webp&s=1109e69db807b153064b1f5b61d22cf1e9fbca05 Another user posted the benchmarks for Qwen 3.8 27B today, and while I think Qwen 27B is a really powerful model, I can't…
Using llama.cpp I seem to be unable to get my to GPUs working tougether correclty, so I need help somehow. Setup: 96GB RAM, one Blackwell 5000 (48GB) and one 3090 (24GB).
Context: Earlier, open-source large models like Mimo V2.5 Pro, DeepSeek V4 Pro (first version), and Kimi K2.5 used to struggle with this prompt, and Qwen 3.6 27B couldn't even render the globe properly. But now, Qwen 3.8 27B is so much bet…
Just hoping to get some datapoints on what is possible if you have 2x 128gb MacBook pros or Studio Ultras.
DeepSeek-V4-Flash is 165 GB on disk, so it does not fit in 64 GB of memory. It still runs, because the model only uses a small part of its weights for each token.
DeepSeek V4 Flash performance on a single MI300X What a single GPU actually delivers when the clients are autonomous coding agents doing real work, not synthetic benchmark load. Built on the open-source deepseek-v4-flash-mi300x serving sta…
DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀 🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge.
So, after weeks of trying to put my finger on it, i decided to finally do a comparative test. I had a long task, a very detailed plan for a major partial refactor and integration of a new system.