Curiosity · AI Model

Reka Flash 3

Reka Flash 3 is Reka AI's 2025 open-weight 21B-parameter reasoning model, released under Apache 2.0. Positioned below Reka Core in size but tuned for strong reasoning and tool use, Flash 3 targets enterprise deployments where open-weight, cost-efficient inference matters — a rare frontier-adjacent team putting serious weight into open models.

Model specs

Vendor
Reka AI
Family
Reka
Released
2025-03
Context window
32,000 tokens
Modalities
text, code
Input price
$0.4/M tok
Output price
$0.8/M tok
Pricing as of
2026-04-20

Strengths

  • Apache-2.0 licensed — truly open commercial use
  • Strong reasoning-per-parameter ratio
  • Hosted on multiple inference providers (Reka, DeepInfra, Together)

Limitations

  • Smaller context window than Reka Core
  • Text-only (no vision in Flash 3 release)
  • Less benchmark coverage than big-lab peers

Use cases

  • Self-hosted reasoning agents on private clusters
  • Cost-efficient chat at scale via Reka API or DeepInfra
  • Enterprise VPC deployments requiring open-weight licence
  • Fine-tuning base for domain-specific reasoning

Benchmarks

BenchmarkScoreAs of
MMLU-Pro~65%2026-04
GPQA Diamond~30%2026-04
HumanEval~82%2026-04

Frequently asked questions

What is Reka Flash 3?

Reka Flash 3 is Reka AI's 21-billion-parameter open-weight reasoning model, released in 2025 under Apache 2.0. It is tuned for cost-efficient reasoning and agent workflows.

How does Reka Flash 3 compare to Reka Core?

Reka Core is a closed multimodal frontier model with 128k context. Reka Flash 3 is smaller, open-weight, text-only, and optimised for reasoning throughput — a cheaper workhorse.

Sources

  1. Reka Flash 3 on HuggingFace — accessed 2026-04-20
  2. Reka AI blog — accessed 2026-04-20