Supersonic Labs logo

Julia 2 Flash

The compact version of Julia 2. Same decisions, 3.3 Γ— smaller.

Julia 2 β€” monochrome geometric artwork

Julia 2 Flash is the compact version of Julia 2. It reads the same inputs (text, JSON, images, audio and live calls) and answers the same question types (choice, score, noul). It uses the same API. The weights are 658 MB, against 2.14 GB for Julia 2.

At a glance

Julia 2 Julia 2 Flash
Weights 2.14 GB 658 MB (βˆ’69 %)
Text only (text.safetensors) β€” 427 MB
GPU memory, peak 3.74 GB 1.08 GB
One text question (GPU) 6.2 ms 5.7 ms
64 text questions, batched (GPU) 295 ms 171 ms
One image question, 1.6 MP (GPU) 93.5 ms 80.3 ms
Load time (GPU) 6.4 s 1.2 s
One text question, Apple M4 CPU (4 threads) β€” 0.3–0.6 s (depends on other load), 0.9 GB RAM

GPU numbers: one NVIDIA RTX PRO 6000, median of 50 runs (Flash: median of three processes).

Quality against Julia 2

All numbers are paired differences against Julia 2 (bf16), in accuracy points, with 95 % bootstrap intervals. Zero means "same as Julia 2".

Test Size Flash βˆ’ Julia 2
SLCMB decisions, dev split 445 cases βˆ’0.09 [βˆ’0.77, +0.56]
SLCMB decisions, held-out split 222 cases βˆ’0.07 [βˆ’0.58, +0.38]
Speech intents (MInDS-14, audio in) 2,413 clips +2.20 [+1.08, +3.36] ΒΉ
Fine listening (FLEURS pt-BR and en-US: which of four sentences, one word apart) 1,566 clips +0.77 [βˆ’0.70, +2.17]
Other languages (MASSIVE, 14 languages not in pt/en/es/de/fr) 4,200 items βˆ’0.33 [βˆ’0.69, +0.02]
Main languages (MASSIVE, pt/en/es/de/fr) 900 items βˆ’0.44 [βˆ’1.00, +0.00]
Images (Imagenette and Oxford Flowers validation) 4,945 photos +0.12 [βˆ’0.34, +0.57]
Release tests (text, image, audio, live) 4 / 4 pass

ΒΉ Read this as "no loss", not as a gain. The 4-bit audio tower moves the intent answers. With the audio tower in full precision the difference is close to zero.

Known trade-offs

  • Small changes per answer. On average the results are the same as Julia 2, but single answers can change. On images, Flash picks the same option as Julia 2 in 96 % of the photos.
  • Rare languages. Over 14 less common languages, Flash is about 0.3 points below Julia 2. The largest single drops are Arabic (βˆ’1.7 points) and Thai (βˆ’1.3 points). If you work mainly in one of these languages, test Flash on your data first.
  • Fine-grained images. On fine-grained photos (102 flower species) the answers move more than on everyday objects (agreement 88 % against 98 %). The mean accuracy is the same.

Usage

uv pip install torch safetensors tokenizers numpy pillow soundfile scipy
hf download SupersonicLabs/Julia-2-Flash --local-dir Julia-2-Flash
import sys; sys.path.insert(0, "Julia-2-Flash/runtime")
from julia2 import Runtime

rt = Runtime("Julia-2-Flash")                               # loads text, image and audio
rt = Runtime("Julia-2-Flash", modalities=("text",))         # text only: opens only text.safetensors (427 MB)

rt.predict_questions(state, {
    "move": {"type": "choice", "instructions": "...", "criteria": {"a": "...", "b": "..."}},
    "risk": {"type": "noul", "instructions": "Is this risky?"},
})

The full API (images, audio, batches and live calls) is the same as Julia 2. Refer to runtime/README.md.

For the lowest memory on CPU, set these variables before you load the model:

export JULIA2_CPU_KEEP_QUANT=1 JULIA2_COMPILED_DEQUANT=1 JULIA2_TOKEN_BUDGET=1024

Files

File Contents Size
text.safetensors decision trunk and token table 427 MB
image.safetensors vision tower 116 MB
audio.safetensors audio tower 115 MB
runtime/ the julia2 runtime
tokenizer/, trunk/, config.json tokenizer and configuration

Download only the files for the inputs you use. The runtime opens only the files for the loaded modalities.

What is inside

  • Trunk weights in int4 (groups of 128), with 8 trunk layers in int8. 20 of the 24 trunk layers are kept.
  • Token table in int4 for frequent tokens and int2 for rare tokens.
  • Vision tower in int8. Audio tower in int4 (groups of 128).
  • Norms, readout and recurrent state stay in higher precision, as in Julia 2.

Licence

Apache-2.0, as Julia 2. Refer to LICENSE, NOTICE, NOTICE_audio.md and licenses/.

Downloads last month
17
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for SupersonicLabs/Julia-2-Flash

Finetuned
(2)
this model

Collection including SupersonicLabs/Julia-2-Flash