C.W.K.
Stream
Lesson 05 of 05 · published

Tradeoffs and the v8 / RWKV-X Future

~12 min · limitations, rwkv-8, rwkv-x

Level 0Observer
0 XP0/50 lessons0/14 achievements
0/100 XP to next level100 XP to go0% complete

Recall — the same wall, in a different shape

RWKV-7 maintains good performance up to about 28K tokens on passkey-retrieval tests. Past that, performance degrades. The shape of the failure mode differs from Mamba's (RWKV's exponential decay produces a different pattern of forgetting than Mamba's selective compression), but the category of failure is the same: finite state can't store arbitrary amounts of information.

This is the recurring theme of this quest. Every architecture that tries to replace attention runs into the same wall at some context length, just at different lengths and with different failure shapes. RWKV's wall is at ~28K. Mamba's is similar. Hyena's is different in shape but present. Transformers don't have the wall, but they have the cost.

Ecosystem reality check

RWKV's ecosystem is impressively broad for an alternative architecture, but it remains much smaller than Transformer's. Fine-tuning recipes, instruct-tuned variants, evaluation harnesses, RLHF infrastructure — all are thinner. Benchmark competitiveness still requires more training tokens to compensate for the architectural constraints, which means small fine-tuning runs hit harder ceilings than equivalent Transformer fine-tunes.

The upside: a smaller ecosystem means more opportunity to contribute. If you're a researcher or hobbyist who wants to leave a mark, RWKV's tooling has more low-hanging fruit than the Transformer ecosystem in 2026.

What's coming: v8 Heron and RWKV-X

RWKV-8 "Heron" introduces two new ideas: DeepEmbed (an edge-MoE-style sparse mechanism without VRAM overhead, allowing larger effective model capacity at fixed inference memory) and ROSA (a neurosymbolic suffix automaton for structured knowledge — bringing explicit symbolic state into a recurrent architecture).

RWKV-X (April 2025) is the family's hybrid response to the recall limitation: RWKV layers + sparse attention layers, mirroring the SSM-attention hybrid pattern from Jamba and Nemotron-H. The convergence to hybrids is happening across families, not just within Mamba.

External links

Exercise

Run a passkey-retrieval test on RWKV-7 yourself: place a known string at varying depths in a synthetic context (10K, 28K, 64K tokens of filler), and ask the model to repeat it. Plot recall accuracy vs. depth. You should see a clear cliff somewhere around 25–32K tokens. That cliff is real and is the most important practical limit to internalize about RWKV.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.