Mathematics + code
Work through multistep problems, generate code and follow structured instructions.
Models / EdgeRock-4BSmall model.
Small model.
Longer horizons.
Reason across text and images with a fixed-capacity history. EdgeRock is our first open 4B World-State model.
01 / CapabilitiesUseful reasoning.
Useful reasoning.
Bounded history.
82.9% on AIME. 4B parameters. One model for text and vision.
EdgeRock retains 96.3% of its source model’s AIME 2024–2026 mean score after replacing full attention with bounded memory. Performance varies by task; the report includes the gains and the tradeoffs.
| Evaluation | EdgeRock-4B | Qwen3.5-4B source | What it probes |
|---|---|---|---|
| AIME 2024–2026 mean | 82.9 | 86.1 | Competition mathematics |
| MATH-500 | 97.5 | 98.0 | Mathematical problem solving |
| MMLU-Pro | 79.8 | 77.7 | Broad knowledge and reasoning |
| SuperGPQA | 60.0 | 52.8 | Specialist knowledge |
| GPQA-Diamond | 64.6 | 77.3 | Graduate science questions |
| LiveCodeBench v6 | 61.6 | 75.7 | Code generation |
| IFEval | 88.0 | 89.6 | Instruction following |
AIME uses mean pass@1 over the three annual sets. Evaluation settings differ across benchmarks. These are results reported in the technical report, rather than a new evaluation conducted by this website. See full protocols and results, §§5–6.
Visual understanding
84.9% mean answer-level accuracy across six reported visual benchmarks, retaining 95.3% of the source average.
Learned retrieval
Memory anchors are selected and updated by the model, rather than supplied as an external document search service.
02 / ArchitectureHistory has
History has
a new home.
A compact state that keeps moving forward.
EdgeRock combines 24 recurrent layers and 8 RMA attention layers. Each attention layer uses a fixed local window and a fixed anchor bank instead of a cache that grows with every token.
How RMA works- Model size
- 4B parameters
- Source architecture
- Qwen3.5-4B
- Recurrent / RMA layers
- 24 / 8
- Local attention window
- 1,024 tokens
- Memory anchors
- 1,024 per KV group
- Commit interval
- 256 tokens
- BF16 attention KV
- 72 MiB per sequence
Cache capacity is configuration-derived and excludes model weights, recurrent state, temporary activations and runtime buffers. At 256K context, Qwen3.5-4B’s corresponding full-attention KV is 8 GiB: a 99.1% reduction for EdgeRock.
03 / SystemsMeasure the
Measure the
whole workload.
Fixed history changes the long-generation memory budget. It also makes larger batches practical. Single-request decoding and batch throughput answer different deployment questions.
| Workload | EdgeRock | Qwen3.5-4B |
|---|---|---|
| 32K input + 32K output End-to-end output tokens/s | 12,124 Batch 512 | 1,889 Batch 64 |
| 128K input + 128K output Decode tokens/s, batch 1 | 233.7 | 151.3 |
Report-era protocol: BF16 weights and attention KV; EdgeRock used FP32 recurrent state at batch 1 and FP16 at larger batches. Batch timing includes prefill and uses prompt sharing; decode-only timing excludes prefill. These results do not represent the newer BF16-state, independent-prompt protocol. Loading, warmup and tokenization are excluded. Full hardware, timing and comparison conditions in §9.
Choose your target device, context length and concurrency before evaluating. Quantized and mobile releases have their own validation scope.
Choose a runtime04 / DownloadsOne architecture.
One architecture.
Your deployment path.
Start with the original checkpoint. Explore smaller packages when your memory budget calls for them.
Quantized adaptations currently target text inference. Custom RMA support is required; a file format alone does not make a standard runtime compatible.
| Release | Use it for | Validation scope | Download |
|---|---|---|---|
| EdgeRock-4B / BF16 | Reference model and research | Text + vision report evaluations | Hugging Face ↗ |
| NF4 / FP8 / INT8 | Smaller PyTorch checkpoints | Functional and limited capability screening | NF4 ↗ · FP8 ↗ · INT8 ↗ |
| GGUF | Custom llama.cpp / Ollama RMA path | Linux x86 CUDA; custom backend required | GGUF release ↗ |
| MLX 4-bit / group 32 | MLX RMA development | Linux CUDA; Apple Metal not validated | MLX release ↗ |
| ONNX 4-bit / group 32 | Custom ONNX CUDA inference | Linux CUDA functional validation | ONNX CUDA ↗ |
| ONNX CPU 4-bit | CPU execution experiments | CPU adaptation; mobile ONNX not validated | ONNX CPU ↗ |
| LiteRT-LM INT4 | Android GPU development preview | Selected Adreno device; custom runtime | Android preview ↗ |
Format adaptations are not equivalent to full benchmark validation. Review each release’s model card, dependencies and license before use. Code is provided under Apache 2.0 with model-specific terms in MODEL_LICENSE.
05 / Model notesA few useful
A few useful
distinctions.
Does constant memory mean unlimited perfect recall?
No. The history state has a fixed capacity. The model learns what to keep and retrieve, so retention is selective. Context length, task difficulty and the chosen runtime still affect useful performance.
Do all downloads support images?
The original research model supports text and vision. The quantized runtime adaptations listed here currently validate text inference; their release notes define the supported inputs.
Can I drop the model into any existing runtime?
The architecture needs RMA attention and anchor updates. Use the supplied code or an adaptation with explicit EdgeRock support. Standard GGUF, Ollama or MLX support alone is insufficient.
Where should I ask about commercial use?
Read the model license and contact jihs@stonehill-tech.com for licensing questions.
Put EdgeRock to work.
Bring a workload, a device and a question. We can start there.