Model
Model explorer

CodeGen 16B (Mono)

OPEN
Salesforce · CodeGen family · released Mar 25, 2022

Early open program-synthesis LLM (350M-16B) trained on TPU-v4; Mono variant was competitive with OpenAI Codex on HumanEval

ReasoningCodingVisionFunction callingTool useAgentic
-26.4
Elo · unrated
Parameters
16.1B
Active params
16.1B (dense)
Context
2K tokens
Architecture
Dense autoregressive transformer (GPT-2-style), 16.1B params, trained on TPU-v4
License
BSD-3-Clause
Languages
—
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
HumanEvalCoding29.3%#158
best: Claude Opus 4.5 · 99.4%
MBPPCoding35.3%#90
best: Llama-3.3-Nemotron-Super-49B v1 (Reasoning On) · 91.3%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
32.2 GB
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
Permissive
QLoRA12.1 GB1× RTX 4070 Ti 16GB
LoRA35.5 GB1× A100 80GB
Full fine-tune259.6 GB2× H200 141GB
QLoRA SFT on ~10k samples ≈ $5.89 (1× RTX 4070 Ti 16GB)
Plan a fine-tune →
CodeGen family
Elo progression across releases
CodeGen 16B (Mono)Mar 2022-26.4CodeGen2.5 7BJul 202397.4
Resources
HFSalesforce/codegen-16B-monoGHsalesforce/CodeGen
API price weights · each benchmark row carries its own source badge (see methodology)