Ternary LLMs making a comeback? Multiple 1.58-bit models released, some beat full-precision baselines
Individual-Dot5488 · reddit · 2026-08-16
Reddit users discuss the resurgence of ternary (1.58-bit) LLMs. In the past month, several labs released ternary models: PrismML's 27B (benchmarks inconsistent), Deepgrove's maple-20b-a1b (runs at 100 tok/s on iPhone), and Doses AI's medical pestle-27b-ternary (outperforms medgemma-27b on medical benchmarks at 8x smaller size). Common issue: long-horizon agentic coding, but users attribute this to lack of RL optimization in new labs, which they plan to address. Hope for a competitive ternary MoE model against Qwen3.8.
More from Models
- Five rounds of adversarial prompting make Opus 5.5 and Argon converge on identical output — LanceCharger · 2026-10-16
- DIY benchmark ranks 18 models on subjective judgments, GLM-5.3 tops consensus match — arkuto · 2026-10-16
- Bindu Reddy: Google's Argon excels at complex doc writing at half the cost of Opus — bindureddy · 2026-10-16
- MAI-2 is 2x bigger than V4-Pro while V4.1F is 3x smaller: Microsoft's scaling ladder questioned — teortaxesTex · 2026-10-16
- Step 5 Preview hits DeepInfra day-zero: 600B MoE, 1M context, $1/M input tokens — StepFun_ai · 2026-10-16
- Liquid AI's d1-ultra Decision Model Hits OpenRouter at $0.042/M Input Tokens — JosephJacks_ · 2026-10-16