Model
Model explorer

Starling-7B (alpha)

OPEN
UC Berkeley (Berkeley NEST) / Nexusflow · Starling family · released Nov 20, 2023

RLAIF fine-tune of OpenChat-3.5/Mistral-7B using the GPT-4-labeled Nectar dataset and APA policy optimization; scored 8.09 MT-Bench (GPT-4 judge) and 91.99 AlpacaEval, best 7B at release behind only GPT-4/GPT-4-Turbo.

ReasoningCodingVisionFunction callingTool useAgentic
390.3
Elo · rank #334
Parameters
7B
Active params
7B (dense)
Context
8K tokens
Architecture
Dense transformer (Mistral-7B via OpenChat-3.5 base)
License
Apache-2.0 (with condition not to compete with OpenAI)
Languages
—
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AlpacaEvalHuman preference92.0%#2
best: Tülu 2+DPO 70B · 95.1%
ARC-ChallengeReasoning63.8%#73
best: Llama 3.1 405B · 96.9%
Arena EloHuman preference1083#48
best: Claude Fable 5 · 1505
BIG-Bench HardReasoning44.4%#106
best: ERNIE 4.5 300B-A47B · 94.3%
GPQA DiamondReasoning29.7%#265
best: GPT-5.6 · 94.6%
GSM8KMath62.4%#111
best: Llama 3.1 405B · 96.8%
HellaSwagReasoning84.9%#42
best: Claude 3 Opus · 95.4%
IFEvalReasoning54.8%#128
best: Gemma 4 26B A4B · 98.5%
MMLU-ProKnowledge31.7%#163
best: Claude Fable 5 · 91.5%
MMLUKnowledge63.9%#183
best: OpenAI o3 · 92.9%
MT-BenchHuman preference8.09#45
best: Hunyuan-Large (A52B) · 9.4
TruthfulQAKnowledge46.4%#63
best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%
WinoGrandeReasoning80.6%#35
best: PaLM 2 · 90.9%
Run it locally
VRAM @ Q4
4.5 GB
VRAM @ FP16
14 GB
Fits on (Q4)
RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
~135 tok/s on RTX 4090 (Q4, llama.cpp)
Quantizations
GGUF Q4 · AWQ
Starling family
Elo progression across releases
Starling-7B (alpha)Nov 2023390.3
Resources
HFberkeley-nest/Starling-LM-7B-alphaDOCSstarling.cs.berkeley.edu
API price weights · each benchmark row carries its own source badge (see methodology)