qwen3-4b-int8

A Core AI bundle of Qwen/Qwen3-4B for the Aether SDK (iOS and macOS 27+).

• Source: Qwen/Qwen3-4B at revision 1cfa9a7208912126459214e8b04321603b3df60c, licence Apache-2.0. The licence is included as LICENSE.
• Changes from the source: converted from PyTorch to Core AI (.aimodel) by Aether forge (recipe qwen3-4b-int8@1). Weights are int8-linear-perchannel (8-bit weights). The tokenizer files are the source's own.

Variants

Variant	Platform	Arch	Compute	Compiled	Assets	Download
macos-any-gpu	macos	any	gpu	no (specialized on first load)	qwen3_4b_int8.aimodel 4.03 GB	4.04 GB
ios-any-gpu	ios	any	gpu	no (specialized on first load)	qwen3_4b_int8.aimodel 4.03 GB	4.04 GB
ios-h18p-gpu	ios	h18p	gpu	yes	qwen3_4b_int8.h18p.aimodelc 4.03 GB	4.04 GB

ios-any-gpu needs the com.apple.developer.kernel.increased-memory-limit entitlement. ios-h18p-gpu needs the com.apple.developer.kernel.increased-memory-limit entitlement.

Verification

Every row is a record in verification/ about exactly these bytes (matched by bundle digest). Reference rows are strict T2 passes of the unquantized export on the same fixture, in verification/reference/.

Variant	Tier	Result	Detail	Device	OS build	Compute	Record
ios-any-gpu	T0	pass		iPhone18,2	24A446	target	c2d4b19c
ios-any-gpu	T2	pass	20/20 strict; profile quantized-8bit; fixture 5ea7a2e85ce97881	iPhone18,2	24A446	target	a5e13d85
ios-any-gpu	T3	pass	copy-fidelity-v1; 100.0% vs reference 100.0%; 50 items	iPhone18,2	24A446	target	16a6dffe
ios-h18p-gpu	T0	pass		iPhone18,2	24A446	target	1bc4eff4
ios-h18p-gpu	T2	pass	20/20 strict; profile quantized-8bit; fixture 5ea7a2e85ce97881	iPhone18,2	24A446	target	2e3efff5
ios-h18p-gpu	T3	pass	copy-fidelity-v1; 100.0% vs reference 100.0%; 50 items	iPhone18,2	24A446	target	de525ebc
macos-any-gpu	T0	pass		Mac17,6	26A434	target	45d0a3c5
macos-any-gpu	T1	pass		Mac17,6	26A434	target	5eec69f7
macos-any-gpu	T2	pass	20/20 strict; 20/20 strict; profile quantized-8bit; fixture 5ea7a2e85ce97881	Mac17,6	26A434	target	2456ea1c
macos-any-gpu	T3	pass	copy-fidelity-v1; 100.0% vs reference 100.0%; 50 items	Mac17,6	26A434	target	a72445b5
unquantized reference (not published)	T2	pass	20/20 strict; 20/20 strict; profile strict; fixture 5ea7a2e85ce97881	Mac17,6	26A434	target	c361bfea

The quantized profile also requires: T2 strict on the unquantized reference export (met by the reference row).

Use

aether run qwen3-4b-int8 --prompt "Hello"

import Aether

let aether = try Aether()
let chat = try await aether.chat("qwen3-4b-int8")
let reply = try await chat.respond(to: "Hello")
print(reply.text)