Model
Model explorer

Reka Flash (21B)

CLOSED
Reka AI · Reka family · released Feb 12, 2024

Original 21B Reka Flash, announced Feb 12, 2024 (documented together with Edge/Core in Reka's April 2024 tech report); iterated into a Sep 2024 build and then superseded by Flash 3/3.1, and no longer separately served on Reka's API under its original form.

ReasoningCodingVisionFunction callingTool useAgentic
633.5
Elo · rank #276
Parameters
21B
Active params
21B (dense)
Context
128K tokens
Architecture
21B dense multimodal transformer (Core/Flash/Edge series)
License
Proprietary
Languages
API price (in/out)
No hosted API
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
GPQA DiamondReasoning34.0%#246
best: GPT-5.6 · 94.6%
GSM8KMath85.8%#63
best: Llama 3.1 405B · 96.8%
HumanEvalCoding72.0%#78
best: Claude Opus 4.5 · 99.4%
MATH-500Math29.6%#166
best: GPT-5 · 99.4%
MMLUKnowledge75.9%#121
best: OpenAI o3 · 92.9%
MMMUVision53.3%#88
best: Claude Fable 5 · 89.3%
MT-BenchHuman preference8.2#41
best: Hunyuan-Large (A52B) · 9.4
PubMedQAKnowledge69.0%#9
best: Palmyra Med 70B · 79.6%
VQAv2Vision78.4%#11
best: Reka Edge (2026) · 88.4%
XWinogradReasoning84.0%#4
best: Sarashina2-70B · 91.8%
Run it locally
Closed weights — available via API only. No local deployment.
Input / M tok
Output / M tok
API price · each benchmark row carries its own source badge (see methodology)