Reka is an AI company that builds natively multimodal models, and its current API offers two of its own models, Reka Edge 2603 and Reka Flash 3, alongside a Reka Vision video API. This hub lists every Reka model named in Reka's public documentation, Hugging Face model cards and news posts as of 17 November 2026, with API IDs, release dates, context windows, prices, licences, availability and the benchmark figures Reka reports. It covers Reka Edge 2603, Reka Flash 3.1, Reka Flash 3, the 2024 Reka Core, Flash and Edge generation, Reka Speech and the Reka Vision API. It is written for developers, buyers, architects and anyone who needs to know which Reka model is which. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from Reka's own documentation, news posts, technical report or official Hugging Face cards, cited in the sources list.
Last checked: 2026-11-17.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Reka's public API documentation as of November 2026. Model details change, so confirm against Reka's current documentation.
Key takeaways
- Two Reka models are on the current models page. Reka's models page lists reka-edge-2603 and reka-flash-3; the same API also serves third-party open models such as DeepSeek, GLM, Gemma and Qwen variants.
- Context windows are 16K and 64K. Reka lists 16K context and 16,384 maximum output for Reka Edge 2603, and 64K context and 65,536 maximum output for Reka Flash 3.
- Prices start at $0.10 per million tokens. The model catalogue lists Reka Edge 2603 at $0.10 input and $0.10 output, and Reka Flash 3 at $0.10 input and $0.20 output, in US dollars.
- Weights are published for Edge and Flash. Reka Flash 3 and 3.1 cards name Apache-2.0; the Reka Edge 2603 card names its own licence, which permits commercial use below $1 million yearly revenue.
- No retirement dates are published. Reka's docs say the model list changes as models are added and retired, but no dated deprecation notice was found on the pages checked.
What is Reka?
Reka describes Spark, Edge, Flash and Core as its natively multimodal models for video, image, audio and text. Its products include Reka Cloud, for training, customising and deploying models on Reka infrastructure, and Reka Vision, a video platform for tagging, searching, reasoning over and clipping video through an API, MCP or the Reka app. Research is published under Reka Labs. The API is OpenAI-compatible, with the base URL https://api.reka.ai/v1 and Bearer-token authentication using a key created in the developer console.
This page covers Reka's own models. Third-party models served through Reka's API are listed in Reka's catalogue but are out of scope here. Spark is named on Reka's home page without further detail on the pages checked, so it has no row. Facts are as of 17 November 2026.
Current Reka models at a glance
As of 2026-11-17. Source: Reka models page, model catalogue and Hugging Face cards. Models are listed by release date, newest first.
| Model | API ID | Category | Context | Licence | Status |
|---|---|---|---|---|---|
| Reka Edge 2603 | reka-edge-2603 | Edge (multimodal, edge deployment) | 16K | reka-edge-2603-license | Current on the API |
| Reka Flash 3.1 | reka-flash-3.1 (as announced in July 2025) | Flash (reasoning) | Not stated by Reka on the pages checked (2026-11-17) | Apache-2.0 | Weights on Hugging Face; not listed on the current models page |
| Reka Flash 3 | reka-flash-3 | Flash (reasoning) | 64K | Apache-2.0 | Current on the API |
| Reka Core (2024) | reka-core, reka-core-20240415, reka-core-20240501 | Core (multimodal) | Not stated by Reka on the pages checked (2026-11-17) | Not stated by Reka on the pages checked (2026-11-17) | Named on the legacy available-models page; not listed on the current models page |
| Reka Flash (2024) | reka-flash, reka-flash-20240226 | Flash (multimodal) | Not stated by Reka on the pages checked (2026-11-17) | Not stated by Reka on the pages checked (2026-11-17) | Named on the legacy available-models page; not listed on the current models page |
| Reka Edge (2024) | reka-edge, reka-edge-20240208 | Edge (multimodal) | Not stated by Reka on the pages checked (2026-11-17) | Not stated by Reka on the pages checked (2026-11-17) | Named on the legacy available-models page; not listed on the current models page |
"Not listed" records what the current models page shows; Reka has not published a dated retirement notice for these models on the pages checked. See Open-Weight vs Closed Models Explained.
Key facts for the main Reka models
As of 2026-11-17. Identical fields for every model; "Not stated by Reka" means not stated by Reka on the pages checked (2026-11-17).
| Model | Released | Max output | Knowledge cutoff | Input / output | Open weights | Parameters | Positioning (provider wording) |
|---|---|---|---|---|---|---|---|
| Reka Edge 2603 | 2026-03-11 (announcement) | 16,384 tokens | Not stated by Reka | Text, image, video / text | Yes | 7B (657M ConvNeXt V2 vision encoder and 6.4B transformer backbone) | "Frontier-level edge intelligence for physical AI" |
| Reka Flash 3.1 | 2025-07-10 | Not stated by Reka | Not stated by Reka | Text / text | Yes | 21B | Update to Flash 3, "particularly strong on coding" |
| Reka Flash 3 | 2025-03-10 (announcement) | 65,536 tokens | Not stated by Reka | Text / text | Yes | 21B | Reasoning model trained from scratch |
| Reka Core (2024) | 2024-04-15 (announcement) | Not stated by Reka | Not stated by Reka | Text, image, video, audio / text | Not stated by Reka | Not stated by Reka | Largest and most capable model of the 2024 series |
| Reka Flash (2024) | 2024-02-12 (announcement) | Not stated by Reka | Not stated by Reka | Text, image, video, audio / text | Not stated by Reka | Not stated by Reka | "Efficient and capable multimodal language models" |
| Reka Edge (2024) | Not stated by Reka (dated API ID 20240208) | Not stated by Reka | Not stated by Reka | Text, image, video, audio / text | Not stated by Reka | Not stated by Reka | Compact model of the 2024 series |
The 2024 modalities come from the abstract of the Reka Core, Flash and Edge technical report, which says the models process text, images, video and audio. Release dates marked "announcement" are the dates of Reka's news posts.
Capabilities stated in the current catalogue
| Capability | Reka Edge 2603 | Reka Flash 3 |
|---|---|---|
| Image input | Yes (models page also lists video) | No |
| Tool calling | Yes | No |
| Reasoning | No | Yes |
| Structured outputs | Yes | Yes |
| Logprobs | Yes | Yes |
| Prompt caching | No | No |
| Chat and streaming | Yes | Yes |
As of 2026-11-17. Source: Reka model catalogue and models page.
Reka model families and naming
- Edge: the compact multimodal line. The 2026 model's ID, reka-edge-2603, carries a year and month suffix (March 2026). Reka describes targets including NVIDIA Jetson, Apple Silicon Macs, Linux and Windows PCs and Qualcomm Snapdragon phones and wearables, and object detection prompted with the format "Detect: {expression}".
- Flash: the mid-sized line. The 2024 Flash was multimodal; Flash 3 (March 2025) and Flash 3.1 (July 2025) are 21B text reasoning models with published weights.
- Core: the largest model of the 2024 series, announced on 15 April 2024, with dated IDs reka-core-20240415 and reka-core-20240501.
- Spark: named on Reka's home page among its multimodal models; no specification was found on the pages checked.
- Reka Speech: announced on 16 September 2025 as a speech transcription and translation model with timestamps.
- Reka EdgeQ: announced on 22 September 2026 under the title "Reka EdgeQ, On-Device"; specifications were not checked for this page.
- Reka Quant: Reka's quantisation technology, released with Flash 3.1, producing a 3.5-bit build (Q3_K_S datatype in llama.cpp).
The legacy page lists both undated IDs such as reka-core and dated IDs such as reka-core-20240415. Where a dated ID exists, pinning it keeps behaviour stable across updates.
When was each Reka model released?
As of 2026-11-17. Dates are those of Reka news posts or the arXiv submission.
| Date | Release | Source |
|---|---|---|
| 2026-09-22 | Reka EdgeQ, on-device | Reka news |
| 2026-04-14 | Reka Edge available on OpenRouter | Reka news |
| 2026-03-11 | Reka Edge (reka-edge-2603) | Reka news |
| 2025-09-16 | Reka Speech | Reka news |
| 2025-07-10 | Reka Flash 3.1 and Reka Quant | Reka news |
| 2025-07-08 | Reka Vision | Reka news |
| 2025-03-10 | Reka Flash 3 | Reka news |
| 2024-10-04 | Reka Flash updates (multimodal understanding, reasoning, agent building blocks) | Reka news |
| 2024-04-18 | Technical report: Reka Core, Flash, and Edge | arXiv 2404.12387 |
| 2024-04-15 | Reka Core | Reka news |
| 2024-02-12 | Reka Flash | Reka news |
How much do Reka models cost?
As of 2026-11-17. Source: Reka model catalogue at developer.reka.ai/models and Reka pricing docs. US dollars per million tokens.
| Model | Input | Output | Cached input | Batch |
|---|---|---|---|---|
| Reka Edge 2603 | $0.10 | $0.10 | Not stated by Reka (prompt caching not listed for this model) | Not stated by Reka |
| Reka Flash 3 | $0.10 | $0.20 | Not stated by Reka (prompt caching not listed for this model) | Not stated by Reka |
- Billing model: Chat Completions is billed per token at each model's rate, with no platform fee or markup, according to Reka's pricing page.
- Price fields: GET /v1/models returns prompt, completion and input_cache_read prices as strings in US dollars per token; a prompt value of "0.00000011" equals $0.11 per million input tokens.
- Prepaid credits: credits are added before requests and spend is tracked on the Usage page; volume or committed-use pricing is by contact with Reka.
Reka Vision API pricing
| Item | Endpoint | Price |
|---|---|---|
| Video indexing | /videos/upload | $0.05 per input video minute |
| Video search | /search | $0.005 per search |
| Image upload | /images/upload | $10 per 1M images |
| Image storage | None | $50 per 1M images per month (deleted after 30 days) |
| Image search | /images/search | $0.003 per search |
| QA chat, streaming QA, indexed and quick tagging | /qa/chat, /qa/stream, /qa/indexedtag, /qa/quicktag | $2 per 1M output tokens |
| Clip generation | /v1/clips | $0.04 per input video minute, plus $0.01 for 1080p |
Video storage is free and deleted automatically after 30 days, and the free tier includes 180 minutes of indexed video (Reka Vision pricing page). See LLMjacking Guide for why prepaid credit and spend tracking matter for exposed keys.
Where are Reka models available?
- Reka API: OpenAI-compatible Chat Completions at https://api.reka.ai/v1; GET /v1/models returns the live roster, and Reka notes an account may see more models.
- Playground: Reka Edge is available for hands-on use in the Reka app, and Flash 3.1 was announced for the Playground.
- Hugging Face: the RekaAI organisation publishes reka-edge-2603, reka-flash-3, reka-flash-3.1 and a 3.5-bit reka-flash-3.1-rekaquant-q3_k_s build.
- OpenRouter: Reka announced Reka Edge on OpenRouter on 14 April 2026.
- Local runtimes: the Flash 3.1 card describes a Llama-compatible format for Hugging Face Transformers or vLLM; the Edge card gives vLLM and llama-server examples.
- Regions: not stated by Reka on the pages checked (2026-11-17).
Restricted-access Reka models
No Reka chat model is described as restricted access on the pages checked. Two items are by request: the 3.5-bit Reka Quant build of Reka Edge is "available on request" according to the Edge announcement, and volume or committed-use pricing requires contacting Reka. Reka's models page notes that individual accounts may see models beyond the public list.
What is the context window of Reka models?
Reka lists a 16K context window and 16,384-token maximum output for Reka Edge 2603, and a 64K context window and 65,536-token maximum output for Reka Flash 3 (models page and catalogue, as of 2026-11-17). Context for Flash 3.1 and the 2024 models is not stated by Reka on the pages checked. For images, the Edge announcement states 64 tokens per image tile, and the Edge card reports 331 input tokens for a 1024 by 1024 image, which leaves room for several images within 16K.
Is any Reka model open source?
Reka publishes weights for three of its models. The Reka Flash 3 and Reka Flash 3.1 cards on Hugging Face carry the Apache-2.0 licence. The Reka Edge 2603 card names a model-specific licence, reka-edge-2603-license, and states that the model can be used commercially if the user makes less than $1 million of revenue a year. Weights for the 2024 Core, Flash and Edge models are not published on the RekaAI Hugging Face organisation. Self-hosting brings its own controls; see the Open-Weight Model Security Guide and Self-Hosted LLMs vs Hosted LLM APIs.
Model cards and system cards
| Document | What it states |
|---|---|
| Reka Edge 2603 card | 7B parameters, BF16; image and video plus text in, text out; reasoning not currently supported; trust_remote_code=True required; float16 on Apple MPS; about 14 GB in float16, 24 GB minimum Mac memory; exported with transformers 4.57.3 |
| Reka Flash 3.1 card | 21B parameters; post-trained with supervised fine-tuning, then reinforcement learning (RLOO) with verifiable rewards; primarily built for English; drop earlier reasoning traces in multi-turn chat |
| Reka Flash 3 card | 21B parameters; trained from scratch; supervised fine-tuning then RLOO with model-based and rule-based rewards; not the strongest choice for knowledge-intensive tasks, pair with web search; often reasons in English; no extensive alignment or persona training |
| Technical report (arXiv 2404.12387) | Reka Core, Flash and Edge process text, images, video and audio; parameter counts not given in the abstract |
Reka does not publish a separate system card on the pages checked (2026-11-17).
Benchmarks
As of 2026-11-17. All figures are provider-reported. No independent leaderboard figures are included.
| Model | Benchmark | Provider-reported result | Source |
|---|---|---|---|
| Reka Edge 2603 | VQA-V2 | 88.40 | Hugging Face card |
| Reka Edge 2603 | MLVU | 74.30 | Hugging Face card |
| Reka Edge 2603 | MMVU | 71.68 | Hugging Face card |
| Reka Edge 2603 | RefCOCO-A / RefCOCO-B | 93.13 / 86.70 | Hugging Face card |
| Reka Edge 2603 | VideoHallucer | 59.57 | Hugging Face card |
| Reka Edge 2603 | Mobile Actions | 88.40 | Hugging Face card |
| Reka Edge 2603 | Latency | Time to first token 0.522 s; mean end-to-end 4.69 s; 5.46 images per second at 1024 by 1024 | Card and announcement |
| Reka Flash 3.1 | LiveCodeBench v5 (full set) | 10 points above Reka Flash 3 | Hugging Face card and news post |
| Reka Flash 3.1 (3.5-bit) | Quantisation degradation | Average degradation 1.6 with Reka Quant, against 6.8 for a standard Q3_K_S routine | News post |
The Edge card compares results with other open models of similar size and notes that some comparison models were measured through an API while others ran locally. Reka also reports that 4-bit quantisation of Edge cuts memory from 13GB to 5GB and keeps over 98% of performance. See the How to Read LLM Benchmarks page for how to read provider-reported scores.
Deprecations and retirements
Reka's models page states that "the model list changes as models are added and retired". A legacy available-models page still names reka-core, reka-flash and reka-edge with their dated 2024 snapshots, while the current models page and catalogue list only reka-edge-2603 and reka-flash-3 among Reka's own models. Reka Flash 3.1 was announced with the API ID reka-flash-3.1 in July 2025 but is not on the current models page. No dated deprecation or shutdown notice was found on the pages checked (2026-11-17). Treat GET /v1/models as the authority and test before relying on any ID not in its response.
Data handling and security
- Authentication: API keys are created in the developer console and sent as a Bearer token in the Authorization header.
- Training use of API data: not stated by Reka on the pages checked (2026-11-17).
- Retention: Reka Vision deletes stored videos and images after 30 days; chat retention is not stated by Reka on the pages checked (2026-11-17).
- Prepaid billing: credits must be added before requests, which caps loss from a leaked key at the credit balance.
- Self-hosted use: running published weights keeps prompts and outputs in the operator's environment; the Edge card requires trust_remote_code=True, so pin and review the repository revision.
- Third-party models: the same Reka API key reaches third-party models in Reka's catalogue, so model allow-lists belong in the calling application or gateway.
See AI Data Retention and Deletion and the On-Device AI Security Guide for edge deployments.
How to choose a Reka model: factual criteria
- Modality: Edge 2603 accepts image and video with text; Flash 3 accepts text only.
- Reasoning: Flash 3 is a reasoning model; the Edge card states it does not currently support reasoning.
- Tool use: the catalogue lists tool calling for Edge 2603 and not for Flash 3.
- Context: 16K for Edge 2603 and 64K for Flash 3.
- Price: both list $0.10 input; output is $0.10 for Edge 2603 and $0.20 for Flash 3.
- Hardware for self-hosting: Edge needs about 14 GB in float16 or about 5GB at 4-bit; Flash has 21B parameters.
- Licence: Apache-2.0 for Flash 3 and 3.1; the Edge licence has a $1 million yearly revenue threshold for commercial use.
- Language: both Flash cards state the models are primarily built for English.
Practitioner checklist
- Call GET /v1/models and record the exact model IDs your account can use in the AI inventory.
- Replace any reka-core, reka-flash or reka-edge 2024 IDs with a model on the current models page and test the output.
- For commercial use of Reka Edge weights, check revenue against the licence threshold and keep the licence text with the deployment.
- Pin the Hugging Face revision and review custom code before loading Edge with trust_remote_code=True.
- Download weights only from the official RekaAI Hugging Face organisation and track provenance.
- Confirm training-use and retention terms in Reka's contract before sending regulated data to the API.
- Store Reka API keys in a secrets manager, keep credit balances low, and watch the Usage page. See the Model Aggregator and LLM Router Key Security page where a gateway or OpenRouter is used.
Standards and references
- Reka: Models (accessed 2026-11-17)
- Reka: Model catalogue and prices (accessed 2026-11-17)
- Reka: API pricing (accessed 2026-11-17)
- Reka: Vision API pricing (accessed 2026-11-17)
- Reka: Reka Edge 2603 model card (accessed 2026-11-17)
- Reka: Reka Flash 3.1 model card (accessed 2026-11-17)
- Reka: Reka Flash 3 model card (accessed 2026-11-17)
- Reka: Reka Edge announcement (accessed 2026-11-17)
- Reka: Reka Flash 3.1 and Reka Quant (accessed 2026-11-17)
- Reka: Reka Core, Flash, and Edge technical report (accessed 2026-11-17)
- Reka: Available models, legacy list (accessed 2026-11-17)
- Reka: News (accessed 2026-11-17)
Related NHI Mgmt Group resources: Sakana AI Models, On-Device AI and AI PCs Hub, Self-Hosted LLM Security Guide, LLM Security Guide.