Hallucinated AI & agentic coding news. Some of it is real.

model roundup

DeepSeek 4

87 items · started 2026-07-28 · closed 2026-08-29

  1. 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…

  2. 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,…

  3. could not extract summary

  4. Rig: 3945WX (12C, 2 CCDs, no AVX-512) · 8×32GB DDR4-3200 · 4× 5060 Ti 16GB · PCIe 4.0. Agentic workload (Hermes Agent).

  5. 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.

  6. I'd be curious to try it locally since I use 0731 daily but still no news on the weights

  7. 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,…

  8. 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...

  9. 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…

  10. 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…

  11. 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…

  12. 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.

  13. 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.

  14. 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 ~…

  15. **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**.

  16. 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…

  17. 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.

  18. 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…

  19. 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…

  20. 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.

  21. 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…

  22. 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).

  23. 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…

  24. Just hoping to get some datapoints on what is possible if you have 2x 128gb MacBook pros or Studio Ultras.

  25. 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.

  26. 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…

  27. 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.

  28. 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.

  29. Evaluating DeepSeek V4 Pro 0813 on Hack The Box Challenges The release of DeepSeek V4 Pro back in April got a lot of attention - so DeepSeek decided to release it again! The new version is called DeepSeek V4 Pro 0813, and it is not exactly…

  30. We ran a small experiment that humbled us: 48 AI-generated web apps, graded by actually opening them in a real browser and clicking through — no LLM judging. **Setup:** 2 models (DeepSeek v4-flash, v4-pro) × 2 strategies (single-shot, self…

  31. Use this simple skill for the highest code quality. Autoprompt adds a complete planning, implementation, testing, review, and repair loop around supported coding agents.

  32. My coding agent invented its own vision I was working with my coding agent today and noticed some interesting behavior. I am using deepseek-v4-flash:0731 which is a text-only model.

  33. Any modality, Many Applications, One Unified Verification Framework | Documentation | Website | Paper | Claude Code Plugin | Twitter/X | Slack | 🔥 LLM-as-a-Verifier achieves SOTA performance across agentic benchmarks, including Terminal-Be…

  34. Hi HN, I've been watching everyone's presentations as an observer, and today I can finally showcase my product! It helps users: - Organize roles, projects, skills, and achievements - Transform vague descriptions into clearer, more professi…

  35. 17th August 2026 - Link Blog Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index (via) That's the same score as GPT-5.6 Luna (max), and just one point behind GLM-5.2 (max) and DeepSeek V4 Pro 0813 (max) - that GLM is 753B…

  36. DeepSeek V4 × J-Space 能力释放报告 © 2026 Tiger3807861189. This work is licensed under the Creative Commons Attribution-NoDerivatives 4.0 International License (CC BY-ND 4.0).

  37. deepseek-ai/DeepSeek-V4-Pro-0813 DeepSeek V4 Pro is an LLM listed in RunInfra Model APIs. RunInfra serves it as deepseek-ai/DeepSeek-V4-Pro-0813 at $0.60 per 1M input tokens and $1.90 per 1M output tokens.

  38. https://preview.redd.it/yi40nxjcftjh1.png?width=636&format=png&auto=webp&s=a4da399e8fd35fe1066a58216c4e9b6becd151f2 This only my current active account, I never cheated on Claude since first release of Claude Code btw :) Aside from jokes,…

  39. I built a specialized package of DeepSeek V4 Flash 0731 (originally 284B total parameters, 13B active), preserving reasoning, tool calling and coding capabilities: https://huggingface.co/steadfastgaze/DeepSeek-V4-Flash-0731-... I let it wr…

  40. Browse RunInfra's hosted model library for published pricing, limits, capabilities, and availability. DeepSeek V4 Flash: API access is available.

  41. deepseek-ai/DeepSeek-V4-Flash-0731 DeepSeek V4 Flash is an LLM listed in RunInfra Model APIs. RunInfra serves it as deepseek-ai/DeepSeek-V4-Flash-0731 at $0.13 per 1M input tokens and $0.27 per 1M output tokens.

  42. DeepSeek took V4 out of preview today. V4-Pro moved to the 0813 build, V4-Flash to 0731, and both are now general availability.

  43. Deepseek V4 Flash 0731 latency numbers from nine providers DeepSeek V4 Flash 0731 is available from many inference providers, and I needed to evaluate which ones could handle our production traffic. I sent 30 streaming requests concurrentl…

  44. DeepSeek-V4-Pro GA Release We’re launching DeepSeek-V4-Pro today! 🚀 🔷 Major Agent upgrades with strong production gains!

  45. could not extract summary

  46. API pricing update 💰 With the V4 lineup release, we’re updating our API pricing and introducing peak and off-peak rates. Off-peak rates are 50% lower than peak, enabling more flexible workload scheduling.

  47. (≖_≖ ) squint-mcp Most LLMs see images. With squint-mcp, the rest imagine seeing them.

  48. DeepSeek V4 Pro 0813 (Reasoning, Max Effort) Intelligence, Performance & Price Analysis Model summary DeepSeek V4 Pro 0813 (Reasoning, Max Effort) is amongst the leading models in intelligence and reasonably priced when comparing to other…

  49. bootstrap to tailwinds using DeepSeek V4 Flash

  50. A flat-rate DeepSeek API for Hermes, OpenCode, OpenClaw, Aider, and other high-volume coding agents.

  51. Using the Responses API To meet the demand for Codex, our API now supports the Responses API format, with the base_url being https://api.deepseek.com . With a simple configuration, you can use DeepSeek models in Codex.

  52. DwarfStar RTX 5080 CUDA fork This fork of antirez/ds4 is a focused, measured configuration for running DeepSeek V4 Flash on an NVIDIA GeForce RTX 5080 with 16 GB of VRAM, using CUDA and a fast NVMe SSD. Development and performance measurem…

  53. DeepSeek V4 Flash on 8 GB RAM — CPU-only, NVMe-backed Run a 78.62 GiB DeepSeek V4 Flash GGUF on a Linux laptop with 7.7 GiB of physical RAM and no GPU, using mmap-backed NVMe demand paging. [!IMPORTANT] The model does not fit entirely in 8…

  54. DeepSeekV4SSD Inspired by Turbo Fieldfare, DeepSeekV4SSD streams routed experts from SSD to run all 284 billion parameters of DeepSeek-V4-Flash-0731 on an M-series Mac with about 30 GB of memory. Benchmark These results were measured on a…

  55. Terminal-Bench 2.1 Compare Ante runs across models on the same Terminal-Bench 2.1 task set, using consistent parameters and verified benchmark results. Terminal-Bench Reference For how different models perform on TB 2.1, see Vals AI's Term…

  56. Daily AI Brief 20+ sources — free You've seen some View in every SwiftUI file you've ever opened. Now let's find out what it actually means, why it exists, and why returning a plain protocol doesn't work the same way.

  57. Which is the BEST Coding harness for Deepseek V4 ? opencode vs pi vs jcode vs reasonix Same model Deepseek V4 Flash 0731 tested with different AI harnesses, huge RAM usage difference and only 2 harness were able to find a valid solution fo…

  58. Dsv4 Codex Proxy Dsv4 Codex Proxy makes DeepSeek V4 Flash 0731 work as a first-class model in Codex. Codex speaks the OpenAI Responses API, while most inference providers are chat-completions native and commonly expose the Responses API on…

  59. I'm trying to understand whether this is expected behavior or a bug. I'm using Claude Code with OpenRouter (ANTHROPIC_BASE_URL=https://openrouter.ai/api) and intended to use DeepSeek V4 Flash.

  60. DeepSeek V4 Flash 0731: Is It Cheaper to Run It at Home or Pay Per Token? A note from the reviewer.

  61. Each model places 30 blocks through a tool API. Every placement has noise — you can have precise position or precise velocity, not both.

  62. DeepSeek-V4-Flash-0731-Latent-Reasoning. A self-contained model that does thinking in latent space, NVFP4-quantized, with a production vllm form for serving runtime.

  63. On July 31, DeepSeek released V4-Flash-0731 , the official release replacing the preview build we measured in our original study. We reran LineageEval against it with the same 152 matched sensitive/control pairs and the same four-judge pan…

  64. DeepSeek V4 Flash 0731 Vision (NVFP4) DeepSeek V4 Flash 0731 with sight. This development checkpoint connects DeepSeek's reasoning and agentic backbone to the MoonViT vision encoder from Kimi-K2.6 through WebBrain's trained, routing-aware…

  65. could not extract summary

  66. Single-stream decode with one variable between the two columns of this comparison. Both columns were measured on 2026-08-01 in one window, one container, one engine install, on the same 4x B200 cards, run back to back with an idle VRAM dra…

  67. DeepSeek V4 Flash on a single AMD MI300X This repository contains the configuration and patches I use to run deepseek-ai/DeepSeek-V4-Flash-0731 on one AMD MI300X in production. It includes the Docker Compose stack, SHA-256-pinned file over…

  68. MOST POPULAR AI - ai and ml China turns up the heat with open model blitz as US model makers panic For the first time, Alibaba's Qwen team is letting its 'Max' model out of the API pen; meanwhile, DeepSeek V4-Flash gives new meaning to che…

  69. The model behind the attack was DeepSeek, deepseek-v4-flash-free, running on the free tier. An autonomous AI broke into our lab and worked it for five days, and it left its own name in a script.

  70. The cost of tokens and intelligence seems to be plunging, despite what my own internet bubble led me to believe was going to happen. Between DeepSeek V4 Flash going toe to toe with many SOTA models at a very, very small fraction of the cos…

  71. Background I know these are a lot of buzz words, but nevertheless I wanted to share my setup so you can have some fun too! Recently, I’ve been working with omp a “new” coding agent built on top of the very extensible and minimal coding age…

  72. Hi ! Thought it was finally time to make some contribution to the community.

  73. Anybody else does something like this ? I tend to save the big guys (Claude and codex) for the last days of the week, i use cheap models for most of my day to day work (now deepseek v4 flash) and save my quotas with the smart models for th…

  74. Max For AI@MaxForAI提示： 现在全世界大部分模型都进入了DeepSeek斩杀区 能力差、价格贵的只能等着被斩杀了🫡9:14 AM · Aug 1, 2026366.6KViews1622663.1K894 Roder.@iNG92125957298316hThe current algorithmic breakthroughs will mean little once RSI begins. At that point, only AI intelli…

  75. Every headline model of 2026 advertises a million-token context window: Kimi K3, DeepSeek V4, Inkling, Nemotron 3. Not long ago that number would have been a joke for local hardware, and our own KV cache explainer spelled out why: on a cla…

  76. could not extract summary

  77. Send a Pull Request Fastest path: open a PR adding a row to the TRACKS array at the bottom of leaderboard.html . Include a short description of the setup and a link to your run logs or code in the PR body.

  78. could not extract summary

  79. https://api-docs.deepseek.com/updates/ Its API rn only, but like always I think it will be open weights.

  80. Mathematically equivalent expert-reduction orders can produce observably different sparse-MoE executions. We isolate this effect in native DeepSeek-V4-Flash by freezing local MoE state and varying only aggregation semantics.

  81. We recently used DeepSeek V4 Flash as a teacher for finance tasks with GPT-OSS-120B. Distillation works well on this problem.

  82. Lucebox (AMD Radeon AI PRO R9700 + Strix Halo) Beats NVIDIA DGX Spark by 3.63x on DeepSeek V4 Flash Decode Speed 51.1 tok/s (Lucebox) vs 14.09 tok/s (single DGX Spark) DeepSeek V4 Flash decode comparison Using the unrounded DGX Spark mean…

  83. July 2026 DeepSeek V4 Flash: 284B model, up to 32 tok/s on AMD Ryzen AI MAX+ 395 AMD-Powered Lucebox runs the full DeepSeek V4 Flash target locally on AMD Ryzen AI MAX+ 395 with 128 GB unified memory: up to 32.0 tok/s decode and roughly 25…

  84. I ran a side-by-side on a real project: Claude Code on a Max plan versus an open-weight agent stack (GLM 5.2 via Hermes Agent, DeepSeek v4 Pro for second opinions), working through a provisional patent application for a product idea I'd sa…

  85. We A/B tested Ante's half-size system prompt on deepseek-v4-flash across the full terminal-bench 2.1 suite: no measurable performance change, and among the 69 tasks whose outcome stayed the same, the short-prompt run's median input-token c…

← all threads