Finished file size
Byte count of the file we ship. Measured, not a class guess.
Products we make well.
Shop lists only units we have bit-identical codec proof for, or that
already serve on AhanaBot with serve_aarm.
You buy the finished file size. Compression is lossless of the packed source.
~1-bit class is a quant size, then AQ1B packs that quant losslessly.
Dense shrink of 1/16–1/2 is the quant step — not a claim that F16 AQ1B is 16×.
Serve compressed: serve_aarm on an AARMv1 face. Never llama. Never GGUF expand for inference.
Lossless AhanaZip
AhanaAI / AhanaZip compression. Bit-identical to the packed source. Serve the finished file — no secret recipe on this page.
Serve face
AARMv1 is what serve_aarm loads. AQ1B is the lossless weight ship. We sell both when we make them well.
Lossless AQ1B — sha-verified
Finished sizes we measured. AhanaZip / AhanaAI lossless of the packed source.
TINY · F16 recoverable · AQ1B
LOSSLESS OF F16
0.6B student F16
1.091 GiB finished
Dense F16 1.198 GiB → 1.091 GiB finished. AhanaZip lossless of F16.
TINY · Q4/Q2 pack · AQ1B
LOSSLESS OF THAT QUANT
0.6B student Q-pack
0.428 GiB finished
Quant source 0.434 GiB → 0.428 GiB finished. Smallest shop file.
MID · Q2 pack · AQ1B
LOSSLESS OF Q2
1.7B reasoner Q2
0.844 GiB finished
Quant source 0.853 GiB → 0.844 GiB finished. Near 1-bit-class size.
HELM · Q4 pack · AQ1B
LOSSLESS OF Q4
4B Q4_K_M pack
2.483 GiB finished
Q4 source 2.497 GiB → 2.483 GiB finished. Lossless of Q4, not of F16.
SPECIALTY · Q2 · AQ1B
LOSSLESS OF Q2 · DATED FILE
4B specialty Q2
1.748 GiB finished
Quant source 1.764 GiB → 1.748 GiB finished. Dated proven file only.
HELM · F16 recoverable · AQ1B
LOSSLESS OF F16
4B helm F16
7.358 GiB finished
Dense F16 8.051 GiB → 7.358 GiB finished. AhanaZip lossless of F16.
AARMv1 — serve_aarm / AhanaBot
These faces already load on our custom server. Chat is live on the 0.6B and 4B helm seats.
TINY · AARMv1 · serve_aarm
LIVE ON AHANABOT LADDER
0.6B serve face
0.825 GiB finished
Dense F16-class ~1.20 GiB → 0.825 GiB finished. AhanaZip / AhanaAI. Live on AhanaBot.
HELM · AARMv1 · serve_aarm
LIVE CHAT · MIRACLE SIZE
4B helm serve face
0.893 GiB finished
Dense F16 8.05 GiB → 0.893 GiB (~9×). AhanaZip / AhanaAI. Serve compressed.
MID · AARMv1
CONTAINER OPENS
1.7B mid face
2.406 GiB finished
Dense F16-class ~3.4 GiB → 2.406 GiB finished. One-file serve container.
CUSTOM · AQ1B then AARM wrap
WE MAKE THIS WELL
Bring an LLM
finished size = packed bytes
Lossless AQ1B of the source you give. Quant 1/16–1/2 if you want 1-bit-class size. Then Jeremiah wraps AARMv1 for serve_aarm.
Order a proven SKU Chat / assemble on AhanaBot
Not listed: 27B, 8B class guesses, Dual_Q1 as “lossless F16”, 9B/Gemma until sha+serve pass. Experts live on AhanaBot. Folder codec: AhanaZip.