Model
Model explorer

PLaMo-100B

OPEN
Preferred Networks · PLaMo family · released Oct 10, 2024

Ground-up 100B trained from scratch by Preferred Elements (a PFN subsidiary) on 2T EN/JA tokens; base for the later PLaMo Prime line.

ReasoningCodingVisionFunction callingTool useAgentic
234.4
Elo · rank #359
Parameters
100B
Active params
100B (dense)
Context
4K tokens
Architecture
Ground-up dense causal decoder-only transformer with QK-normalization and Z-loss training stabilization
License
Preferred Elements dual license (separate commercial and non-commercial terms)
Languages
2+
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
HumanEval+Coding22.0%#37
best: Mistral Small 3.2 (24B) · 92.9%
JASTERKnowledge77.5%#1
best: this model · 77.5%
MMLUKnowledge60.3%#204
best: OpenAI o3 · 92.9%
MT-BenchHuman preference7.7#51
best: Hunyuan-Large (A52B) · 9.4
RakudaHuman preference9.73#1
best: this model · 9.73
Run it locally
VRAM @ Q4
—
VRAM @ FP16
200 GB
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
PLaMo family
Elo progression across releases
PLaMo-100BOct 2024234.4
Resources
HFpfnet/plamo-100bDOCStech.preferred.jpAPIAvailable via Preferred Networks API
API price weights · each benchmark row carries its own source badge (see methodology)