Reka drops Flash 3.1 with near-lossless 3.5-bit compression
Reka just open-sourced Flash 3.1, a 21B model with major coding upgrades, alongside its new Reka Quant library. The team managed to squeeze the model down to 3.5 bits for llama.cpp with minimal accuracy loss, giving developers a compact agent base that punches well above its weight.
Reka Flash 3.1 is a compact 21-billion-parameter base model built specifically for coding and agentic fine-tuning. Alongside the model, Reka dropped Reka Quant, a new open-source library that squeezes the weights down to 3.5 bits with almost zero impact on accuracy.
Everything is available now on Hugging Face and GitHub for developers who want to run capable agents without renting expensive cluster hardware.
Heavyweight coding in a 21B footprint
Flash 3.1 brings substantial improvements in logic and coding over its predecessor, Flash 3. On the LiveCodeBench v5 benchmark, it scores 10 points higher than the previous version.
Reka reports that on coding-related benchmarks, the 21B model punches well above its weight class, going toe-to-toe with heavier weights like Qwen3-32B and OpenAI's o3-mini.
The secret sauce is a major overhaul of Reka's reinforcement learning pipeline. The team introduced a brand-new RL algorithm paired with significant scalability upgrades, designed to sharpen multi-step execution for autonomous agents.
Squeezing llama.cpp down to 3.5 bits
Even at 21B, serving full-precision models locally can stretch smaller GPUs. That is where Reka Quant comes in.
Sub-4-bit compression usually wrecks model performance. Standard quantization pipelines converting to llama.cpp's Q3_K_S format typically result in an average performance drop of 6.8 points.
Reka Quant circumvents this by using self-distillation, calibrated error reduction, and fast distributed proxy Hessian computations for fast LDLQ. Squeezing Flash 3.1 down to a 3.5-bit Q3_K_S file results in just a 1.6-point drop, making the low-bit compression practically lossless.
The library exports directly to standard llama.cpp data types like Q3_K and Q4_K. Reka actually dogfoods this internally, using customized quantized Flash 3.1 models to orchestrate its web-browsing agent Reka Research and multimodal platform Reka Vision.
How to get it
The weights for Flash 3.1, its 3.5-bit release, and the Reka Quant repository are free to grab on Hugging Face and GitHub.
If you prefer managed inference, the model is also live on the Reka Playground and available via API under the identifier reka-flash-3.1.
Why it matters
If you are building local agent pipelines or tinkering on constrained hardware, this release hits a rare sweet spot. A 21B parameter count already sits in a manageable tier, and getting near-lossless 3.5-bit quantization means you can run an o3-mini-tier coding orchestrator on consumer hardware.
We think open-sourcing the quantization library alongside the model is the real win here. Instead of forcing teams to pick between high latency, huge VRAM bills, or degraded quantizations, Reka delivers a ready-to-run recipe for efficient local agents.
Sources (1)
- reka.ai(opens in a new tab)Originalreka.ai