DGX Spark

Which models can your machine actually run? Magnitude is a 100% free, open source desktop app that: - profiles your hardware and runs sample calculations - predicts tok/s for every model before you download - recommends the best models, from fast to smart Pick your models and it handles the rest: - downloads and tunes the models for your hardware - connects agents like pi, opencode, hermes in one click - runs models on demand as your agent works Works on whatever hardware you already own: MacBook (Intel or Apple Silicon), Mac Mini, Mac Studio, DGX Spark, Strix Halo, any NVIDIA/AMD GPU, or just a CPU Download it on macOS, Windows, or Linux: Open source:

• magnitude.dev
• github.com

I started to play with Qwen Image 2.1 in Early Access and it's really good! 🔥 I've generated images locally using Diffusers on DGX Spark and MFLUX on M5 Max (working on a PR). Here I've tested Image Conditioning from a single image. Love it!

UPDATE: Qwen3.8-Flash for a single DGX Spark 🔥 - 117 tok/s prose & 180 tok/s code at 8 streams. - Optional official Nvidia NVFP4. - 24/7 auto-restart supervisor. - Cached-token reporting in every response. - Peak memory down from 101 to 91 GiB. - LOTS of bugs were fixed. This is still the BEST model to run on a single spark. Full details below 👇 Get it here:

@jvr0x

Big update to the @Alibaba_Qwen Qwen3.8-Flash-Next single DGX Spark recipe! > 𝗪𝗵𝗮𝘁'𝘀 𝗻𝗲𝘄 🎁 > • Measured on one DGX Spark, 262K context, MTP k=3, aggregate tok/s at 1 / 2 / 4 / 8 streams: > Prose: 38.0 / 61.1 / 89.2 / 117.4 Code: 53.8 / 87.5 / 131.8 / 180.2 > • Long context holds: MTP keeps working at a 185K-token prompt (35.4 tok/s decode), prefill ~2,000 tok/s from 4K to 185K > • ~1M-token KV pool at the full 262K context (FP8 KV) > • NVIDIA's official NVFP4 checkpoint now runs on one Spark, with chat, tool calls and vision w

• github.com

that super good speeds

@ViC305

Qwen3.8-Flash-Next -> 79.95 tok/s. 🔥 ONE DGX Spark. Full 262K cache configured. 🚀 Qwen3.8-Flash-Next EXL3 just got another major update. > New measured default: > MTP ndt=5 DSpark, dc=0.6 8-bit KV 262,144-token cache > At an actual 240K-token prompt: > 72.0 tok/s decode ~1,150 tok/s prefill Exact needle retrieval > 𝗙𝗣𝟭𝟲 𝗞𝗩 → 𝟴-𝗕𝗜𝗧 𝗞𝗩 > 4K context: 66.6 → 69.8 tok/s > 128K: 66.6 → 69.5 tok/s > 240K: 65.1 → 72.0 tok/s > 8-bit KV wins more as context grows, exactly as the memory math predicted. > `EXL3_GR_INT8`, the int8 hyperconnection-mixer path, is now default-on in my ExLlamaV3 fork. PR #3 merged into master at `523ecd3`. > And the real context ceiling is the MODEL, not the Spark. > Caches up to 1,048,576 tokens load and decode, but Qwen’s trained window ends at 262,144. Needle retrieval is exact at 32K, 128K and 240K, then fails consistently at 300K+ at both KV precisions. > One import

Qwen3.8-Flash-Next on one DGX Spark ran an open coding job on a real repo for nearly an hour smooth, stable, no collapse. - 125B MoE. Text, image, and video. - FP8 KV Speculative decoding. - Up to 512k context with YaRN. - TP=1 on one Grace Blackwell box with 128 GB unified memory. -

• github.com

This is so clutch. I have issues with this all the time on my DGX Sparks.

@onusoz

DGX Spark users > Make your agents run your inference engines with OOMwrap > Protect your machine from freezing up if they accidentally launch something that takes too much memory, like a model with wrong concurrency or context settings > Couple this with rfjakob/earlyoom, and your DGX Spark will never freeze again > The difference is that earlyoom is a global watcher, and oomwrap (by me) watches individual processes that are run through it > oomwrap includes a memory-safe-launch skill, when installed, makes the agent use it by default for running inference engines > I will make a video about this very soon! > Repo:

• github.com

148 KB. That’s the entire download for this FPS. No textures. No models. No sound files. No launcher. No install. Everything is generated at runtime from ~4,500 lines of JavaScript. Wave survival, headshots, hitscan, tracers, sprint fatigue. Runs in a browser tab. All locally built on one DGX Spark with Qwen3.8 Flash Next/EXL3

While many are transitioning to DeepSeek v4.1 Flash and other models, @plotarmordev has continued working on important PRs and fixes. It's still the most widely used recipe for 2× DGX Sparks.

@plotarmordev

12 PRs merged on DeepSeek V4 Flash (2x DGX Spark), still the most used recipe, and we're keeping the improvements coming > The update fixes tool-call truncation crashes, tightens startup and benchmark scripts, and makes status checks report failures instead of silently passing 👇

Incredible work! As soon as I can get my hands on a DGX Spark, I'm combining this with my work on @OmarchyMac and omarchy-mlx to bring this to Linux on Apple hardware. Who do I know that has good connections at NVIDIA to make this happen?

@ashxhart

MCDMA 0.1.18 is out ✅ > Larger Registered Buffers Teardown fixes A CLI tool Bug fixes >

• github.com

视频生成加速框架 FastVideo 最新释出 FastH3 8-Step V2，并全面打通 Mac 本地 MLX 推理链。 它不是单纯的模型搬运库，而是一套覆盖分布式微调与端到端优化的完整工作流（目前 4.4k Stars）。对于想要在本地完成高品质视频生成的开发者，这次更新直接命中了算力和硬件门槛的痛点。 核心工程进展： • 算力开销大幅压缩：新发布的 FastH3 8-Step V2 基于 MiniMax-H3 进行 DMD2 步进蒸馏，引入高达 80% 的视频稀疏注意力（Video Sparse Attention），极大降低了推理成本。 • Apple Silicon 原生支持：告别云端依赖。借助 MLX 框架与 FastMetal-QAD，Mac 用户现在可以原生运行从 1.3B 到 14B 参数的视频生成模型。 • 多端适配与实时编辑：除了主流 NVIDIA 显卡，现已支持 DGX Spark 环境（注：ARM64 架构目前暂无预编译 wheel，需从源码编译 CUDA kernel）。其内置的 Dreamverse 模块可实现本地视频流的实时“Vibe Directing”控制。 如果你习惯在 macOS 桌面上（例如 32GB 统一内存的 Mac 环境）进行模型部署与自动化测试，FastVideo 提供的清晰 CLI 与 Python API 能帮你快速跑通从代码到视频的最后一公里。 项目文档与源码：

• github.com

Qwen3.8-Flash-Next EXL3 just got a BIG one-Spark update. 🔥 58.8 tok/s single-stream through native ExLlamaV3.🚀 157.6 tok/s aggregate through vLLM across 8 streams.🤯 FULL 262,144 context on ONE DGX Spark. 4.05 bpw EXL3 pack. A much better serving envelope. 𝗕𝗘𝗙𝗢𝗥𝗘 → 𝗡𝗢𝗪 Previous public headline: 47.6 tok/s greedy p50 64K configured context MTP k=2 Concurrency not characterized Now: Native ExLlamaV3: 58.8 tok/s single stream vLLM + vllm-exl3: ~50–55 tok/s single stream 155.6 tok/s @ 4 streams 157.6 tok/s @ 8 streams Configured context: 64K → 262,144 KV pool: 416,163 tokens at the default 262K config 𝗠𝗧𝗣 𝗞=𝟯 𝗜𝗦 𝗡𝗢𝗪 𝗧𝗛𝗘 𝗪𝗜𝗡𝗡𝗘𝗥 Current controlled sweep @ 4K: No draft: 27.77 tok/s MTP k=2: 47.39 MTP k=3: 50.09 At 32K: No draft: 27.58 k=2: 46.60 k=3: 49.73 So the fixed/current build flips the old result: k=3 is now the sweet spot. There is one important boundary I found: At 163,840 PROMPT tokens, MTP acceptance collapses to zero. Above that point speculation becomes slower than running without a draft

With the last commit into DwarfStar now you can use DeepSeek v4.1 Flash in a single DGX Spark as well, with SSD streaming. Around 9 t/s generation. It works also dual-spark RDMA at ~22 t/s.

Open weights. Shared progress. MiniMax H3 is moving fast. We built MiniMax H3 for video generation with native stereo audio and multimodal reference control. The open-source community is making that capability faster, more accessible, and easier to build on. Recent highlights: • FastH3 — FastVideo, Nuva Lab and NVIDIA: 4-step distillation, now running on DGX Spark and Apple Silicon. • Sol-H3 — NVIDIA’s SANA team: now on DGX Spark with a two-stage H3 + LTX-2.5 pipeline. On 8×B300, the team reports 15 seconds of 768p video + audio in 6.6 seconds of warm inference.* • VDN — Haocheng Xi and the OpenVDN team: rethinking attention for faster H3 inference, with weights, training and inference code released. • PDD — NVIDIA’s distillation method, brought to H3 by Alibaba PAI as 8-step Acc-LoRAs, now supported in ComfyUI. • LightX2V — 4- and 8-step Turbo LoRAs, with workflows for text, image and reference-conditioned video + audio. Behind every release are people training, optimizing, quantizing, testing and sh

Updates to GLM 5.3 Flash EXL3 on 2x DGX Sparks 👇

@plotarmordev

Six fixes just landed for GLM-5.3-Flash EXL3 > Setup is less fragile, custom settings are easier to use, and benchmarks now work against password protected servers. > Your default settings are the same, so nothing changes unless you want it to. Thanks again to all contributors! 👇

2× DGX SPARK OWNERS REJOICE! DeepSeek-V4.1-Flash at 3.30 bpw EXL3, targeting just TWO DGX Sparks. 🔥 And yes, VISION is included. DeepSeek already ships its routed experts in FP4. I pushed that expert bank to a 3.30 bpw using my internal SAGE-EXL3 dynamic quant tool average while preserving the rest of the model, including Vision, Engram conditional memory, and the non-expert/source-format weights that are not part of the 3.30 bpw expert quant. TP4 was step one. Now TP2 is here. 𝗗𝗘𝗘𝗣𝗦𝗘𝗘𝗞-𝗩𝟰.𝟭 𝗙𝗟𝗔𝗦𝗛 𝗢𝗡 𝟮× 𝗦𝗣𝗔𝗥𝗞 SAGE-EXL3: 3.30 bpw routed-expert average Mixed precision: K2 → K8 Full pack: 415.8 GiB 31 shards 40 routed-expert layers Vision included. Engram preserved. Non-expert/source-format weights preserved outside the 3.30 bpw expert-bank average. SAGE is an internal quantization workflow I use so I’m not just applying one flat precision across every expert tensor. That’s about as much as I want to say about the method for now. 😁 𝗧𝗛𝗘 𝗧𝗣𝟮 𝗧𝗔𝗥𝗚𝗘𝗧 The full pack is ~415.8 GiB. Roughly ~189 GiB is the

This is just incredible! 🚀

@pratikg

BREAKING: Qwen 3.8 Flash Next now runs more than twice as fast on an NVIDIA DGX Spark. > 105.5% over baseline, up from 25.9% yesterday morning. The Mac track is at 73.8% and climbing! > Every frontier model is on that board now, including Qwen3.8-Max optimizing the engine that runs Qwen. A model making its own runtime faster 🔁 🤌 > is a challenge on @YukonResearch, where multiplayer autoresearch happens. Many humans, many agents, one hard problem, one open scoreboard. > Qwen 3.8 Flash Next is an open-weight model, so anyone can pull it apart and make it quicker on hardware they already own. > Accelerating open intelligence with open frontier research.

• cuda.fast