Signal 3.8 Flash Next (AgentionAI)
Model card (32👍, 17.2k downloads at capture; license: qwen-community-1.0 — inherited from the Qwen base, worth noting for any non-personal use): a “minimally invasive” fine-tune of Qwen3.8-Flash-Next (177B MoE) for lower latency and better token efficiency. Drop-in replacement — the backbone, vision encoder, and multi-token-prediction draft head all work exactly as in the base.
The pitch, measured:
- MATH-500: 26% fewer reasoning tokens, half as many hesitation markers, higher score (99.2% vs 97.5%)
- Terminal-Bench 2.0 (agentic coding): same tasks in 25% less wall time
- Speculative decoding gets faster on top of the speedup: MTP draft acceptance 69% → 90% — more tokens verified per forward pass; run with
--spec-type draft-mtp
The method claim is the interesting part: “Signal gets there by being direct, not by cutting corners. Reasoning models spend a large share of their tokens on hesitation and re-checking… Signal keeps every step that does work and drops the ones that only describe the process” — preambles, sign-offs, narration dropped; substance kept. And it’s trained by self-distillation: Qwen3.8-Flash-Next’s own answers generated under a be-direct instruction that the released model no longer needs — no external data, no other model’s outputs, base knowledge and voice intact.
GGUF quants at capture: AP-Q4_K_XL, AP-IQ4_XS, Q8_0, plus mmproj-F16 (the vision projector — image-text-to-text pipeline).
Vault kin
- Qwen3.8 finetunes — the family ledger; Signal is the token-efficiency entry
- Qwen model family · Local LLM inference — Flash-Next is the household’s Strata-tier MoE lane (125B-class); Signal slots in as a llama.cpp-ready efficiency variant
- The thesis kin: nerkyor’s Coder390 (fixed over-verification by post-training) and the DGX Spark finding (over-verification, not inability) — Signal is the third capture in the “reasoning waste is fixable by tuning” family, now at base-model scale with self-distillation instead of SFT+RLOO
- llama-swap fit: GGUF drop-in for the existing Flash-Next entry; the MTP+spec-dec numbers matter for the DFlash2/speculative-decode lane