EdgeRock-4B hits 82.9% on AIME with fixed-size memory, runs 512 chats on one GPU
eyishazyer · x · 2026-11-04
- Fixed memory no longer means forgetful: EdgeRock-4B uses fixed-size memory (Rock Memory Anchors) and scores 82.9% on AIME (2024–2026, mean pass@1), close to its growing-memory base Qwen3.5-4B at 86.1% — while the best other bounded-state model tested, RWKV-7 G1j (14B), manages only 18.8%.
- The anchors mechanism is the breakthrough: every 256 tokens the model decides which details earn a spot in a never-growing anchor bank, forcing it to learn what's worth keeping.
- Long-context recall: across 256 contexts of 16K–130K input tokens, distant-memory query accuracy is 94.1%; with all 8 RMA modules disabled it drops to 0.4%.
- Throughput: because memory stays fixed, one Hopper 80GB GPU runs 512 concurrent long chats (32K in + 32K out) at 12,124 tok/s output — 6.42× Qwen3.5-4B (64 chats, 1,889 tok/s).
- Source: Rock Matrix, the EdgeRock technical report (arXiv 2610.10118) and GitHub README.
Related event: EdgeRock-4B Open-Sources Constant-Memory Inference for LLMs(3 posts)→
More from Models
- Sub-4B model agent benchmark: Qwen3.5-4B hits 51.4% strict pass, thinking mode adds up to 20 points — KienShen · 2026-11-04
- Redditor proposes an 'SWD bench' to benchmark LLM decompilation instead of just debating it — ThinkExtension2328 · 2026-11-04
- Axion 1.0: ByteDance-Backed Open Agent Foundation Model With 3-Stage Recipe — Yuqi Jia · 2026-11-04
- Grok now introduces itself with he/him pronouns, sparking 'what did they do to Grok' reactions — djcows · 2026-11-04
- eliebakouch: This Is Also a Great Way to Distill Chinese OSS Models — eliebakouch · 2026-11-04
- Subscriber demands refunds, alleging Claude Code headless mode drains Max plan limits ~1.8x faster — jyongchul · 2026-11-04