Reka Flash 3.1
Release in the Reka Flash family · version reka-flash-3.1
A 21B-parameter language model released by Reka in July 2025 as an update to Reka Flash 3. Reka post-trained it with supervised fine-tuning on synthetic and public datasets followed by large-scale reinforcement learning with verifiable rewards, with a focus on coding and on serving as a base for fine-tuning on agentic tasks. It writes a reasoning trace before its answer.134
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- Model hub: 3.5-bit quantized version (Hugging Face) (external site: huggingface.co)
- Release notes: Release announcement (external site: reka.ai)
- Paper: Reinforcement learning write-up (external site: reka.ai)
Availability and license
Component reuse rights
- weights
- Reviewed qualifying license recorded — check scope and conditions
- code
- Unknown — no complete fact-level rights review
- data
- Unknown — no complete fact-level rights review
- documentation
- Unknown — no complete fact-level rights review
No complete system-rights review is recorded for this release.
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | Full weights are on Hugging Face, plus a separate 3.5-bit quantized version.1 |
| Inference codeIs code for running the model published? | Public | The card documents running the model with Hugging Face Transformers and with vLLM; it is released in a Llama-compatible format usable by libraries that support Llama.1 |
| Training codeIs the code used to train the model published? | Unknown | The card and posts reviewed do not link training code. |
| Training-data informationDoes the information cover provenance, scope, acquisition, selection, labeling, processing, and where data or alternatives can be obtained? Access alone does not establish completeness. | Partial | The card says supervised fine-tuning used synthetic and public datasets; the reinforcement learning post says math prompts came from Numina-1.5 and code prompts were difficult coding problems with test cases. Pretraining data for the Flash 3 base is described only as publicly accessible and synthetic datasets.145 |
| Training-data accessCan the training data be obtained? This is independent of information completeness and reuse rights; original unshareable data need not be downloadable. | Unknown | Not assessed. |
| Complete training pipelineIs the complete base-training and preprocessing pipeline published, including configuration? Fine-tuning code or an inference SDK alone is insufficient. | Unknown | Not assessed. |
| Legacy data assessment (v0.1)Historical assessment combining download access and disclosure. Preserved for traceability; excluded from the v0.2 tier calculation. See the new separate assessments above. | Unknown | Not assessed. |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The reinforcement learning post describes Reka's REINFORCE-based algorithm (dynamic sampling, token-level loss, gradient clipping, handling of long samples, on-policy updates) and how RL and SFT examples were separated, but not full hyperparameters or data mixtures.4 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Partial | The card shows evaluation results as a chart; no evaluation code or prompts are linked.1 |
What it is useful for
Run and use notes
- The card documents the prompt format (human and assistant turns separated by <sep>), says generation should stop at <sep> or <|endoftext|>, notes that the model uses the cl100k_base tokenizer, and recommends dropping earlier reasoning traces in multi-turn conversations.1
- The card says the model is primarily built for English and should be treated as an English-only model, although it can understand other languages to some degree.1
Organization context
Provenance and derivatives
Reka describes Flash 3.1 as an update to its Reka Flash 3, which it says was pretrained from scratch. The Llama-compatible format refers to how the weights are packaged for existing libraries.165
- Derived from: Reka Flash 3 (external site: huggingface.co) — Earlier Reka release; no separate catalog record.
Other releases in the Reka Flash family
No other releases in this family have been assessed.
What this catalog does not know
- Training code: unknown.
- Training-data access: unknown.
- Complete training pipeline: unknown.
- Legacy data assessment (v0.1): unknown.
Have a primary source? How to report a correction.
U.S. eligibility
Eligible · basis: U.S. headquarters
The model is published by Reka in its verified RekaAI organization on Hugging Face. Reka AI, Inc. states in its privacy policy that it is headquartered in the United States, and its June 2026 announcement describes it as headquartered in San Francisco (see the Reka record). Reka says the Flash 3 base was pretrained from scratch.17895
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.