Reka Flash 3
Reka's 21B open-weight reasoning model with budget-forced thinking, built for on-device deployment.
About
Reka Flash 3 is Reka AI's 21-billion-parameter open-weight reasoning model, released March 10, 2025. It thinks inside <reasoning> tags before answering, and Reka's signature trick is budget forcing: you can cut the thinking short at a fixed budget and still get a reasonable answer, trading depth for latency on demand. The model is English-first with a 32K context window, and at 21B parameters it compresses to about 11GB in 4-bit quantization for local deployment.
Third-party tracking via Artificial Analysis (reported by dataconomy) puts it at 66.9% MMLU-Pro, 52.9% GPQA, 89.3% MATH 500, and 43.5% LiveCodeBench — respectable for its size though behind larger reasoning models. Reka's own blog shows AIME 2024 accuracy scaling with reasoning budget. Weights are Apache 2.0 on Hugging Face, served through Reka's API and OpenRouter, and a multimodal successor, Flash 3.1, arrived in July 2025.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- MMLU-Pro: 66.9%, GPQA: 52.9%, MATH 500: 89.3%, LiveCodeBench: 43.5%, AIME 2025: 33.7% (third-party Artificial Analysis data via dataconomy.com)
- AIME 2024 (cons@16): Reka reports accuracy scaling with reasoning budget via budget forcing (provider-reported, reka.ai blog)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: real Reka AI release Mar 10, 2025 with Apache-2.0 weights on Hugging Face; provider technical blog with the budget-forcing mechanism; third-party Artificial Analysis scores via dataconomy. Open weights. Strengths: novel controllable-reasoning design, permissive license, honest English-only scoping. Docked for: (1) third-party scores are mid-pack for its class and vendor's own numbers are self-reported; (2) English-first focus limits multilingual use.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Reka AI release 2025-03-10; 21B, 32K context, budget-forcing reasoning
- · Apache-2.0 weights on HF; API via Reka and OpenRouter
- · Third-party Artificial Analysis: MMLU-Pro 66.9, GPQA 52.9, MATH 500 89.3
- · Mid-pack third-party scores; English-first limits multilingual use