Model

Sarashina2-70B

OPEN

SB Intuitions · Sarashina2 family · released Aug 6, 2024

MIT-licensed Japanese base model trained ground-up on ~2.1T tokens of Japanese, English, and code; not instruction-tuned.

ReasoningCodingVisionFunction callingTool useAgentic

14.5

Elo · rank #401

Parameters

70B

Active params

70B (dense)

Context

8K tokens

Architecture

Llama2-style dense decoder-only transformer

License

MIT

Languages

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

BIG-Bench HardReasoning63.9%#74

best: ERNIE 4.5 300B-A47B · 94.3%

GSM8KMath1.1%#199

best: Llama 3.1 405B · 96.8%

HellaSwagReasoning62.7%#138

best: Claude 3 Opus · 95.4%

HumanEvalCoding28.0%#160

best: Claude Opus 4.5 · 99.4%

MGSMMath48.8%#45

best: OpenAI o4-mini · 93.7%

MMLUKnowledge63.0%#191

best: OpenAI o3 · 92.9%

OpenBookQAReasoning39.2%#46

best: Claude 1 · 90.8%

TriviaQAKnowledge53.7%#40

best: Sarvam-1 (2B) · 90.6%

XWinogradReasoning91.8%#1

best: this model · 91.8%

Run it locally

VRAM @ Q4

40 GB

VRAM @ FP16

140 GB

Fits on (Q4)

M3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB

70B does not fit in a single RTX 4090's 24GB VRAM at Q4 (~40GB needed); no verified single-GPU measurement found.

Quantizations

GGUF Q4

Sarashina2 family

Elo progression across releases

Sarashina2-70BAug 202414.5

Resources

HFsbintuitions/sarashina2-70b

API price weights · each benchmark row carries its own source badge (see methodology)