event

AI Benchmark Release

23 items · started 2026-10-22 · closed 2026-11-09

  1. 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…

  2. inspired by Bijan Bowen's video - Subway FPS test. Hardware details: RTX 5090, 256GB DDR4 3200Mhz 8 channel, AMD EPYC 7532 CPU (32 core).

  3. 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…

  4. 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…

  5. This should mean Deepseek 4.1 Flash is excluded

  6. 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…

  7. 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.

  8. 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.

  9. 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.

  10. 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…

  11. Hi HN, I'm Furqan founder of Nebula. We've spent a lot of time building, using and breaking browser agents.

  12. 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.

  13. I just keep waiting for grok 4.8 because i just hate how 4.7 is eating my money, hope they learn from other companies that the newest model is Smarter and cheaper, Also the DeepSeek 4.1 flash on cursor feels more expensive than the api cal…

  14. Local Code Local Code is a private AI agent for Mac, iPad and iPhone, written mostly in Swift. It puts three things in one app: an LLM inference engine for Apple Silicon, a coding harness tuned to each model, and an agent runtime with Pyth…

  15. Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points. Low-bit quantization loses long-tail facts first, and the stan…

  16. As frontier language models grow increasingly capable and autonomous, independent third-party evaluations are critical for characterizing their systemic safety risks. While closed-source releases are routinely accompanied by comprehensive…

  17. Cursor on X: "DeepSeek V4.1 Flash is now available in Cursor! Every effort level sits on the CursorBench cost frontier: from $0.60 to $0.98 per task, no model scores higher.

  18. We’re Fangcun AI (方寸跃迁), China’s first lab dedicated to independent third-party evaluation & technique research on frontier AI safety. Today we’re publishing our evaluation of DeepSeek V4.1 Flash-1017, the first report in our long-term thi…

  19. 24-hour benchmark of 4 AI gateways on real coding-agent sessions with DeepSeek V4.1 Flash: OpenRouter :nitro, Vercel AI Gateway, Ramp Router and Keln on cache hit rate, provider moves, cache after a break, reasoning delivered at max, decod…

  20. I have built an indie hacker project; my main flow was: 1. chatting with Claud Code about the general plan 2.

  21. DeepSeek V4.1 Flash 1017's best Semi-Private scores are 94.5% on v1 at high reasoning and 78.1% on v2 at max reasoning. Pass/fail per reasoning level across each benchmark.

  22. I wanted to stretch my quota by having Claude Code hand mechanical coding work to cheaper models, so I measured it before relying on it. Most of what I expected turned out wrong.

← all threads