PrismML released Ternary Bonsai 2 27B, a ternary-weight model based on Qwen3.8 27B. It is about 5.9 GB—roughly 9x smaller than the full-precision counterpart—while retaining 98.2% of aggregate benchmark performance. PrismML says the release improves agentic coding, multimodal reasoning, and long-horizon tool use, and ships under Apache 2.0.

Key Takeaways

  • ✓The ternary-weight model is about 5.9 GB and 9x smaller than its full-precision counterpart.
  • ✓PrismML reports 98.2% aggregate benchmark retention with gains in agentic coding, multimodal reasoning, and long-horizon tool use.
  • ✓It is available under Apache 2.0 with Hugging Face weights, a paper, a WebGPU demo, and a supporting llama.cpp fork.
⚡

Key Decision Metrics at a Glance

TL;DR Verdict
🏆 Benchmark RankRank
PrismML reports 98.2% aggregate benchmark retention with gains in agentic coding, multimodal reasoning, and long-horizon tool use.
Full Leaderboard→
💰 Cost & PlansCompare
Qwen3.8-27B · in: $0.3 / out: $0.9
Compare Plans↗
🛠️ Ecosystem & IDERules
Cline (VS Code Agent)
Setup IDE Rules↗
🧭

Turn your technical choice into a development budget

Compare 40 dev plans & simulate token costs vs $20/mo subscriptions

⚔️ Compare Model→💰 Compare plans & limits↗
Action HubReady to adopt this in production?

Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.

⚔️ Launch VS Compare→💳 View Plans↗
Up NextReflection Beam weights are live: 501B/23B MoE under Apache 2.0, BF16/FP8/NVFP4 on Hugging Face, 80.1% Terminal-Bench 2.1
Read Now→