reka-flash-3
Reka Flash 3 is a 21B general-purpose reasoning model that was trained from scratch. It was trained in synthetic and public datasets for supervised finetuning, followed by RLOO with model-based and rule-based rewards. It performs competitively with proprietary models such as OpenAI o1-mini, making it a good foundation to build applications that require low latency or on-device deployment. It is currently the best ope...
Params
20.9 B
Context
32,768
Downloads 30d
901
Likes
390
Download history
tracking started — chart appears after 7 days of snapshots (5 recorded)901 downloads in the last 30 days
Can you run it?
Estimated VRAM at 8K context unless noted. Pick your hardware to see the verdict per quantization.
| File | Quant | Size | Est. VRAM | Verdict on RTX 4090 · 24 GB |
|---|---|---|---|---|
| model.safetensors | bf16 | 41.8 GB | 49.6 GB | ❌ Won’t fit |
| model.safetensors (bf16, full) | bf16 + 32K ctx | 41.8 GB | 59.0 GB | ❌ Won’t fit |
Estimate: file size × 1.1 + KV cache at 8K + 0.5 GB overhead. Not a benchmark — how we calculate this.
Run it
copy-paste, exact tags checked against the Hub$ curl -s https://aimodelscomparison.com/api/v1/models/reka-flash-3
{
"hf_id": "RekaAI/reka-flash-3",
"params_b": 20.91,
"context_length": 32768,
"license": { "id": "apache-2.0", "commercial": "yes" },
"downloads_30d": 901,
"vram_estimates": [
{ "quant": "bf16", "gb": 49.6 }
],
"updated_at": "2026-10-19T06:01:13Z"
}
Use via API
via OpenRouter · USD per 1M tokens · synced 2026-10-24$0.10
input / 1M
$0.20
output / 1M
—
cached input / 1M
$0.30
1M in + 1M out
- Context window
- 65,536 tokens
- Max output
- 58,982 tokens
- Providers
- 1
- Cheapest provider
- Reka · $0.10 / $0.20
- Best throughput
- —
- AA Intelligence / Coding / Agentic
- 6.4 / — / —
✗ Tool calling
✓ Reasoning
✓ Structured outputs
✗ Image input
✗ File input
✗ Audio input
✗ Prompt caching
Specifications
- Architecture
- LlamaForCausalLM
- Parameters
- 20.9 B
- Tensor type
- BF16
- Context length
- 32,768
- Vocabulary
- 100,352
- Layers / heads
- 44 / 64
- Licence
- apache-2.0
- First seen on the Hub
- 2025-03-11
- Training datasets
- undisclosed
- Added to our catalog
- 2026-10-20
Compare with any text-generation model