1× Mac
155tok/sEveryday
1 request, llmprobe standard prompt, thinking on, with MTP + PLD (per recipe) Source
Intelligence40with thinking
- Experimental
- License unclear
Local AI · Recipe · 1× RTX 3090
Qwen3.8-Flash-Next EXL3 3.05 bpw with Strata on 1× RTX 3090 (USB dock) + Mac mini M4
Qwen3.8-Flash-Next EXL3 3.05 bpw with Strata on 1× RTX 3090 (USB dock) + Mac mini M4: 33.8 tok/s according to github.com (dataset as of Nov 3, 2026).
by 0xSero
33.8tok/sMixed
1 request, mean over 29 realistic document prompts (3.7–19.8K tokens), with MTP (per recipe)
Source: github.comIntelligence (original model) · Artificial Analysis40with thinking
≤ 3 bit: quantization may cost quality
Engine
- Engine
- Strata v0.1.41 (Niko1221, fb58e0dbc839) + Patches dieses Repos (Engine strata_o37u, Config LS) unter macOS über tgcu-CUDA-Shim auf tinygrads NV-Backend (3090 per USB); kalte Experten per Metal auf dem Mac mini M4 (16 GiB)
- Quantization
- EXL3 3.05 bpw (turboderp, Branch 3.05bpw_h5_ng5; alle 512 Experten, nichts geprunt), Dense-Teil + mmproj IQ3_S GSQ-RCO GGUF, KV k8v4; Experten verteilt auf 3090-VRAM, Mac-mini-RAM (LRU 6.400 Slots) und NVMe
- Model family
- Qwen3.8-Flash-Next
- Context
- 204,800
- Creator
- 0xSero
- GitHub stars
- 7
- Repo updated
- Nov 1, 2026
Measurements
Every sourced value of this recipe, each with its condition and source. Bars relative to the largest value in the group.
Decode · 1 request
Prefill by context
What you need
- Hardware
- 1 × NVIDIA RTX 3090
- Weights
- Main weightsturboderp/Qwen3.8-Flash-Next-exl3 other
- otherISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF Apache-2.0
- MTPQwen/Qwen3.8-Flash-Next other
- Engine
- Strata v0.1.41 (Niko1221, fb58e0dbc839) + Patches dieses Repos (Engine strata_o37u, Config LS) unter macOS über tgcu-CUDA-Shim auf tinygrads NV-Backend (3090 per USB); kalte Experten per Metal auf dem Mac mini M4 (16 GiB)
- Context
- 204,800 tokens
Notes
What matters before you rebuild it.
- ≤ 3 bit: quantization may cost qualityAt 3 bit and below the model may answer noticeably worse than the original. The intelligence number refers to the original.
- Custom kernel required
- vision
- License unclear
Sources
Related recipes
4× RTX 3090
110tok/sEveryday
1 request, prose prompt, with MTP (per recipe) Source
Intelligence40with thinking
- License unclear
2× DGX Spark
89.3tok/sEveryday
1 request, prose prompt, with MTP (per recipe) Source
Intelligenceno independent value
- Custom kernel required
- License unclear
2× DGX Spark
74tok/sEveryday
1 request, prose prompt (long story), thinking off, with MTP Source
Intelligence40with thinking
- Custom kernel required
- License unclear
4× DGX Spark
73tok/sEveryday
1 request, decode at 32K context (speed table, ring of four), thinking on, xhigh Source
Intelligence40with thinking
- Custom kernel required
2× DGX Spark
69.6tok/sEveryday
1 request, prose prompt, with MTP (per recipe) Source
Intelligence40with thinking
- License unclear
Local AI in your company?
In a workshop we work out which models and which hardware fit your tasks, and build the first agent on your infrastructure.