DeepSWE Preview deepswe-preview

agentica-org · deepswe agentic coding

unranked — no quality signals yet

params
32.8B
arch
dense
license
apache-2.0 open weights
released
Jul 2025
downloads/30d
1.6k

# architecture

attn shared GQA 64q/8kv feed-forward (dense) every parameter, every token out ×64 layers 32.8B — no routing
layers
64
d_model
5120
heads
64q / 8kv
head dim
128
vocab
151936
family
qwen3

signal path of one layer, generated from the registry's structured fields — dimension lines quote the model's real numbers; an MoE trace forks at the router, a dense trace runs straight through. Geometry fields come from the repo's config.json.

# why ranked

Overall: provisional — score 44.2 ● low — a single signal; treat with caution

provisional: too few independent signals for a numbered position — the score is shown, but this model sorts after every ranked model.

signalweightinput (0–100)
aa_intelligence 0.70
bench_composite 0.30 44.2

benchmark panel evidence:

complete panelscore (0–100)
swe-bench-verified-v144.2

missing signals are dropped and the remaining weights renormalized — never imputed.

Coding: provisional — score 33.3 ● low — a single signal; treat with caution

provisional: too few independent signals for a numbered position — the score is shown, but this model sorts after every ranked model.

signalweightinput (0–100)
aa_coding 0.30
aider_polyglot 0.30
swe_bench_verified 0.40 33.3

missing signals are dropped and the remaining weights renormalized — never imputed.

Agentic: #15 — score 33.3 ● low — a single signal; treat with caution

signalweightinput (0–100)
swe_bench_verified 0.60 33.3
swe_rebench 0.40

missing signals are dropped and the remaining weights renormalized — never imputed.

full methodology

# trend

OGM score · last 51 days
OGM score over 51 days

methodology changed during this history window; score movement across that boundary is not model movement. See methodology v5.

# benchmarks

benchmarkscoresourcedate
SWE-bench Verified 58.8 SWE-bench Jun 2025

# where to run

no known API providers — download the weights from Hugging Face and run them yourself.

source aliases
swebench
agentica-org/DeepSWE-Preview