Mia’s Favorite
GLM 5.3 Flash
TensorFold · EXL3 4-bit · DFlash2
60.4 tok/s
Decode speed, one stream
- Peak
- 108.8 tok/s ×4 streams
- Prefill
- 1,978.9 tok/s 32k prompt
- Context
- 1M ×4 sharing a ~2.9M pool
21 stars0 forks1 contributor
Last updated
Local AI,
tested in the open.
I run open-weight and frontier models side by side, same prompt, one shot each, and publish every raw output. See what they build, and where they break.
Building with AIPushing Local AI ForwardOpen-Weight ModelsNVIDIA DGX SparksPremium Local AI RecipesQuantization Deep DivesModel Comparisons & RecommendationsBenchmarksNo Cherry-PickingTips & TricksAgentic Workflows
New recipe · my #1 for two Sparks
GLM 5.3 Flash, now on two DGX Sparks.
Powered by TensorFoldA million tokens of context, images and video, and four conversations at once, served from two Sparks on your desk with TensorFold. Free, with every patch and every number published.
- One request
- 60.4tok/s
- Four at once
- 108.8tok/s combined
- Context
- 1Mtokens a request
- Prefill
- 1,978.9tok/s, 32k prompt
Found the needle in a 981,841-token prompt.
See all recipes
github.com/MiaAI-Lab
About
Hi, I'm Mia! I'm publicly developing the frontier of local and secure intelligence.
I'm a developer who got hooked on running AI locally. Most days I'm deploying and optimizing local AI models to run the BEST way possible for NVIDIA DGX Sparks and other devices I have in my lab.
From GLM, DeepSeek, Qwen, and others, I deploy them all, and make them work at their best state possible on the available hardware.
I'm also comparing and benchmarking all closed frontier AI models, including Claude, GPT, Grok and others, to see how they compare to the best local has to offer.
Models offered
The frontier of local intelligence: free, open-source recipes for NVIDIA DGX Sparks and RTX GPUs.
Each recipe runs on my own Sparks, and the benchmarks, patches and launch scripts are all published on GitHub. Clone one and serve a frontier-class model on your own hardware.
- GitHub stars
- 5,607
- Forks
- 789
- Contributors
- 95
Across 19 recipes and sparkDash
Qwen3.8 Flash Next
vLLM · NVFP4
48.7 tok/s
Decode speed, one stream
- Peak
- 163 tok/s ×8 streams
- Prefill
- 1,769 tok/s 32k prompt
- Context
- 262k · 512k with YaRN
576 stars86 forks8 contributors
Last updated
DeepSeek V4 Flash
vLLM · DSpark
38.6 tok/s
Decode speed, one stream
- Peak
- 139 tok/s ×6 streams
- Prefill
- 18.0 s to first token, 32k
- Context
- 1M
1,439 stars196 forks25 contributors
Last updated
See all 19 models
Plus 16 more recipes for DGX Sparks and NVIDIA GPUs, and sparkDash, the dashboard I benchmark with.
Services
Get your local AI lab online and running flawlessly, and optimized for your hardware.
I'll install and tune local AI on your NVIDIA DGX Sparks, the same way I run my own lab, from a single machine to a four-node cluster.
Personal Install by Mia
Pick the setup that matches your hardware. Every install is done personally by me.
- 1×Solo DGX SparkOne machine, installed and tuned for local AIView ↗
- 2×2× DGX Sparks clusterTwo Sparks set up to work as oneView ↗
- 4×4× DGX Sparks clusterFour Sparks for the biggest open modelsView ↗
Booked and paid securely through Whop
What customers say
From verified customers.
Support & contact
How to support the frontier of local intelligence.
Everything here is independent. Donations pay for GPU time, API credits and the coffee behind late-night runs.
Pushing Local AI Forward
Become a sponsor of Mia's AI Lab on GitHubMonthly or one-time, through GitHub Sponsors
Support Mia on Ko‑fiA quick tip toward the lab, through Ko‑fi
or donate with crypto
Donate with crypto through NOWPayments. Pick an amount and a coin, and you'll finish the payment on their secure checkout.
Contact
A question, a model you want tested, a collab idea? I read every message.
or find me on