Most of the AI headlines this year have been about trillion-parameter flagships chasing frontier benchmarks. Reka shipped something almost the opposite on October 9, 2026: a 7-billion-parameter vision-language model built specifically to run on the kind of hardware that doesn't have a data center behind it — a robot's onboard computer, a phone, a laptop with no GPU to spare. The pitch isn't "smartest model." It's "fast and cheap enough to put a camera-equipped AI inside things that aren't servers."
That's a real, narrower claim than a frontier chatbot launch, and it's worth checking against the actual numbers rather than the marketing framing — which is what this piece does, benchmark by benchmark.
Quick summary: Reka released Edge 2603, a 7-billion-parameter vision-language model, on October 9, 2026. It encodes a 1024×1024 image in 331 tokens, scores 88.40 on VQA-V2, 74.30 on MLVU video understanding, and 93.13 on RefCOCO-A object detection, with a reported end-to-end latency of 4.69 seconds per request (Reka separately reports a batched throughput of about 5.46 images per second). Reka's own comparison puts that well ahead of Gemini 3 Pro's 16.67-second latency on the same kind of task, while using roughly a third of the image tokens Gemini 3 Pro or Cosmos-Reason2 8B need. It runs on Jetson Thor, Jetson AGX Orin, and Apple Silicon Macs with 24GB+ RAM, with quantized builds for Jetson Orin Nano, Samsung S25, and iPhone. API pricing is $0.10 per million tokens in and out. The license allows free commercial use only for companies under $1 million in annual revenue — everyone else needs a separate agreement.
What Edge 2603 actually is
Edge 2603 is a 7B-parameter multimodal model — it takes images and text in, text out — built around a specific design goal: spend as few tokens as possible encoding what a camera sees, because every token an edge device has to process costs latency and battery. Reka says it encodes a 1024×1024 image in 331 tokens. For comparison, Nvidia's Cosmos-Reason2 8B needs 1,063 tokens for the same resolution, and Gemini 3 Pro needs 1,094 — meaning Edge 2603 is doing the same visual-encoding job in roughly 30% of the tokens (Source: AI Weekly).
It also ships with tool-calling support and a 16,384-token context window — modest next to frontier models' million-token windows, but enough for the kind of short, repeated perception-and-action loop an edge device actually runs: look at a frame, describe or classify what's in it, decide an action, repeat.
The benchmark numbers
On the tasks Reka published, Edge 2603 scores 88.40 on VQA-V2 (visual question answering), 74.30 on MLVU (long-video understanding), and 93.13 on RefCOCO-A (locating a described object in an image — the skill a robot arm or a phone camera app actually needs). Reported end-to-end latency is 4.69 seconds per request; Reka separately reports a batched throughput of about 5.46 images per second on its test setup (Source: Benchable.ai).
| Model | Tokens per 1024px Image | Reported Latency | Context Window | Price (per 1M tok) |
|---|---|---|---|---|
| Reka Edge 2603 | 331 | 4.69 sec | 16,384 tok | $0.10 in / $0.10 out |
| Cosmos-Reason2 8B | 1,063 | Not published | — | Open weight (self-host) |
| Gemini 3 Pro | 1,094 | 16.67 sec | 1M+ tok | Higher, general-purpose pricing |
A caveat worth stating plainly: these latency and token-efficiency comparisons come from Reka's own launch materials, not an independent third-party test run under identical conditions. Gemini 3 Pro is also a much larger, general-purpose model doing a different job — it isn't optimized for this specific narrow task the way Edge 2603 is, so the comparison favors Reka by design. That doesn't make the numbers false, but it does mean "faster than Gemini 3 Pro" should be read as "faster at this specific narrow vision task," not as an overall capability claim.
Built to run anywhere, including fully offline
Edge 2603 targets Nvidia's Jetson Thor and Jetson AGX Orin boards (the chips inside a lot of today's robots and drones) and Apple Silicon Macs with at least 24GB of RAM for the full model. Reka also ships quantized builds small enough for Jetson Orin Nano, a Samsung Galaxy S25, or an iPhone. A 4-bit quantized version cuts memory from 13GB down to 5GB while reportedly retaining over 98% of the full model's performance — the kind of compression ratio that makes the difference between "needs a workstation" and "runs on the phone in your pocket."
If you're evaluating this for an actual product, test the quantized build on your target hardware before committing — a 98%-performance claim is Reka's own number on its own benchmark suite, and quantization sometimes degrades unevenly on the specific edge cases that matter for your use case (small text in images, cluttered scenes, etc.), not uniformly across the board.
The license has a revenue cap
Edge 2603's weights are open, but the license isn't MIT or Apache — it's a custom Reka license that permits free commercial use only for companies with annual revenue under $1 million. Cross that line and you need a separate commercial agreement with Reka. That's a meaningfully more restrictive stance than, say, Xiaomi's MIT-licensed MiMo-V2.6 family, released about two and a half weeks earlier on September 22, 2026, and it's worth factoring in before building a product roadmap around this model if your company isn't a small startup.
Who this is actually for
- Robotics and drone teams that need visual perception running on Jetson-class hardware without a round trip to the cloud.
- Mobile app developers who want on-device image understanding — object detection, visual search, accessibility description — without per-request cloud latency.
- Early-stage startups under the $1M revenue threshold who want a capable, cheap, commercially-usable vision model without a licensing negotiation.
- Anyone benchmarking small vision-language models against each other for a narrow, speed-sensitive deployment, rather than shopping for a general-purpose chatbot.
01Is Reka Edge 2603 free to use?
The weights are open to download, and commercial use is free if your company's annual revenue is under $1 million. Above that, Reka requires a separate commercial license. Using Reka's own hosted API costs a reported $0.10 per million tokens in and out.
02Is Edge 2603 actually faster than Gemini 3 Pro?
On Reka's own published comparison for this specific image-encoding and latency test, yes — 4.69 seconds versus 16.67 seconds, and about a third of the tokens per image. That's a narrow-task comparison from the model's own maker, not an independent, apples-to-apples benchmark across general capability, so treat it as a strong but vendor-reported claim.
03What hardware can run Edge 2603?
The full model targets Nvidia Jetson Thor, Jetson AGX Orin, and Apple Silicon Macs with 24GB+ RAM. Quantized versions (as small as 5GB in 4-bit) are built for Jetson Orin Nano, Samsung Galaxy S25, and iPhone.
04How does it compare to Xiaomi's MiMo-V2.6, released a few weeks earlier?
They're built for different jobs. MiMo-V2.6 is a much larger, omnimodal model (up to 1.02 trillion parameters) aimed at frontier-class text, image, video, and audio reasoning. Edge 2603 is a 7B model purpose-built to be small, cheap, and fast enough to run on a robot or a phone. They're not really competing for the same use case.
If you're trying to judge whether a specialized small model like Edge 2603 actually beats a general-purpose one for your use case, that's exactly the kind of thing worth testing directly rather than taking a launch blog's word for it — LumiChats gives you 40-plus models, from frontier flagships to leaner specialists, in one place, so you can run your own prompts and images across several of them before deciding what's actually worth building on.
