Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
/**
* The starter list shown with the optional real engine. It is display only: nothing downloads from it (models.install refuses when the real
* engine is on), because these files are named without a hash. Models come from the Models screen of the native engine, which pins them.
* Never relabel these as demo weights.
*/
export const REAL_VARIANTS = [
{
id: 'qwen2.5-0.5b-q4km',
name: 'Starter',
family: 'Qwen2.5-0.5B-Instruct',
quant: 'Q4_K_M',
label: 'Starter · 0.5B Q4_K_M',
params: '0.5B',
blurb: 'Runs on almost anything — the fastest way to see the whole loop',
role: 'option',
tier: 'starter',
ramClass: { minGb: 8, maxGb: 12, label: '8 GB RAM class' },
diskGb: 0.49,
downloadGb: 0.49,
contextMax: 32768,
layerCount: 24,
experts: 0,
kvKbPerToken: 6,
hasDraft: false,
kvModes: ['INT8', 'FP16'],
tags: ['starter', 'smallest'],
minMemoryGb: 8,
source: {
repo: 'Qwen/Qwen2.5-0.5B-Instruct-GGUF',
file: 'qwen2.5-0.5b-instruct-q4_k_m.gguf',
bytes: 491_400_032,
},
},
{
id: 'qwen3-coder-30b-iq1m',
name: 'Coder',
family: 'Qwen3-Coder-30B-A3B-Instruct',
quant: 'UD-IQ1_M',
label: 'Coder · IQ1_M',
params: '30B',
blurb: 'Best when memory is limited and coding is the priority',
role: 'option',
tier: 'coding',
ramClass: { minGb: 24, maxGb: 32, label: '32 GB RAM class' },
diskGb: 9.6,
downloadGb: 9.6,
contextMax: 262144,
layerCount: 48,
experts: 128,
kvKbPerToken: 48,
hasDraft: true,
kvModes: ['INT8', 'FP16'],
tags: ['coding'],
minMemoryGb: 24,
source: {
repo: 'unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF',
file: 'Qwen3-Coder-30B-A3B-Instruct-UD-IQ1_M.gguf',
bytes: 9_630_000_000,
},
},
{
id: 'qwen3-30b-q2',
name: 'Qwen',
family: 'Qwen3-30B-A3B',
quant: 'Q2_K',
label: 'Qwen · Q2_K',
params: '30B',
blurb: 'Fastest full-model choice',
role: 'option',
tier: 'fast',
ramClass: { minGb: 40, maxGb: 48, label: '48 GB RAM class' },
diskGb: 11.3,
downloadGb: 11.3,
contextMax: 131072,
layerCount: 48,
experts: 128,
kvKbPerToken: 48,
hasDraft: true,
kvModes: ['INT8', 'FP16'],
tags: ['fastest'],
minMemoryGb: 40,
source: {
repo: 'unsloth/Qwen3-30B-A3B-GGUF',
file: 'Qwen3-30B-A3B-Q2_K.gguf',
bytes: 11_260_000_000,
},
},
{
id: 'qwen3-30b-iq2m',
name: 'Qwen',
family: 'Qwen3-30B-A3B',
quant: 'UD-IQ2_M',
label: 'Qwen · IQ2_M',
params: '30B',
blurb: 'Balanced quality and speed',
role: 'best-fit',
tier: 'balanced',
ramClass: { minGb: 44, maxGb: 48, label: '48 GB RAM class' },
diskGb: 10.9,
downloadGb: 10.9,
contextMax: 131072,
layerCount: 48,
experts: 128,
kvKbPerToken: 48,
hasDraft: true,
kvModes: ['INT8', 'FP16'],
tags: ['balanced', 'recommended'],
minMemoryGb: 44,
source: {
repo: 'unsloth/Qwen3-30B-A3B-GGUF',
file: 'Qwen3-30B-A3B-UD-IQ2_M.gguf',
bytes: 10_870_000_000,
},
},
{
id: 'qwen3-30b-iq3xxs',
name: 'Qwen',
family: 'Qwen3-30B-A3B',
quant: 'UD-IQ3_XXS',
label: 'Qwen · IQ3_XXS',
params: '30B',
blurb: 'More quality, more memory, slower output',
role: 'option',
tier: 'quality',
ramClass: { minGb: 60, maxGb: 64, label: '60–64 GB RAM class' },
diskGb: 12.9,
downloadGb: 12.9,
contextMax: 131072,
layerCount: 48,
experts: 128,
kvKbPerToken: 48,
hasDraft: true,
kvModes: ['INT8', 'FP16'],
tags: ['quality'],
minMemoryGb: 60,
source: {
repo: 'unsloth/Qwen3-30B-A3B-GGUF',
file: 'Qwen3-30B-A3B-UD-IQ3_XXS.gguf',
bytes: 12_890_000_000,
},
},
];