RWKV 7 'Goose' 1.5B

Positioning

RWKV 7 'Goose' 1.5B is a dense 1.5B-parameter model from the RWKV community, released under the permissive Apache 2.0 license. Its defining feature is a linear-time inference architecture that maintains constant memory usage regardless of context length, enabling a 1,048,576-token context window. This makes it distinct among open-weight models, as most transformers incur quadratic memory costs with long sequences. The small parameter count and efficient architecture position it for edge deployment where memory and latency are constrained.

Strengths

• Extreme context length with constant memory: The linear-time architecture means memory usage does not grow with context length, allowing a 1M-token window that would be impractical for transformer models of similar size.
• Permissive Apache 2.0 license: No restrictions on commercial use, modification, or redistribution, making it suitable for proprietary applications.
• Tiny quantized footprint: At Q4_K_M the model is ~0.8 GB on disk, and even with KV cache overhead it can fit in under 2 GB, enabling deployment on low-power devices.
• Edge-optimized design: With 1.5B parameters and a deployment class of 'edge', this model is purpose-built for scenarios where hardware resources are minimal.

Limitations

• Small parameter count limits raw capability: At 1.5B, the model cannot match the reasoning depth or knowledge breadth of larger models. Operators should expect lower quality on complex tasks.
• No community benchmarks available: We do not have independent measurements of performance on standard tasks. Published vendor metrics should be treated as best-case.
• Architecture maturity: RWKV is a newer architecture compared to transformers; ecosystem support (tooling, fine-tuning libraries) may be less mature.
• KV cache overhead still significant: While the architecture is linear, the KV cache for 1M tokens can still require substantial memory (estimated 30-50% of model size added), which may strain edge hardware.

What it takes to run this locally

At FP16, the model is ~3 GB on disk. Quantized versions reduce this significantly: Q8_0 ~2 GB, Q4_K_M ~0.8 GB, Q2_K ~0.5 GB. For typical use with moderate context (e.g., 128K tokens), add ~30-50% for KV cache and framework overhead, so a Q4_K_M deployment might need ~1.2 GB total. This fits comfortably on a consumer GPU with 4 GB VRAM or even on CPU with sufficient RAM. Deployment class is edge: single GPU with 4-8 GB VRAM, or CPU-only systems.

Should you run this locally?

Yes if you need to process extremely long sequences (up to 1M tokens) on memory-constrained hardware, or if you require a permissive license for commercial edge deployment. No if you need high-quality outputs on complex reasoning tasks, or if you rely on mature ecosystem support and community benchmarks.

Catalog cross-links

• RWKV 7 'Goose' 7B
• Edge deployment guide