PrismML just shrank a 27-billion-parameter AI model down to 5.9 gigabytes. That's small enough for a laptop, and it still keeps 98% of the power of the full-size version.
On September 17, PrismML released its new flagship open-weight model: Ternary Bonsai 2 27B. The headline number is the file size. Built on top of Alibaba's Qwen3.8 27B, the model replaces almost every weight in the network with just three values: -1, 0, or 1. Add a bit of FP16 scaling for each group of weights, and the whole 27.8-billion-parameter model comes out to roughly 1.76 bits per weight. That's a 27B model you can download and run on a laptop.
The old rule of thumb was blunt. A 27B model needs about 54 gigabytes at standard 16-bit precision, and even a 4-bit compressed version still runs close to 18 gigabytes. Ternary Bonsai 2 27B needs 5.9. PrismML says that's more than nine times smaller than the full-precision original. Across a 20-benchmark suite covering reasoning, math, coding, instruction following, vision, and agentic tool use, it scores 83.9 against Qwen3.8's own 85.4. That's a strong showing. Do the math: that's 98.2% of the original's performance, packed into a fraction of the space.
The model also carries a 262,000-token context window, reads images as well as text, and handles coding and multi-step agentic tasks. Nothing about that spec sheet is subtle. PrismML gives it away under an Apache 2.0 license: free to download from Hugging Face, free to use commercially, no fees attached. On an Nvidia RTX 5090, PrismML clocks it at up to 143 tokens per second. On a phone, it's slower. But it runs.
PrismML's first Bonsai, released earlier this year and based on Qwen3.6 27B, made headlines as the first 27B-class model to run on a phone at all. It shipped in two builds: a 3.9-gigabyte, 1-bit version for phones, and a 5.9-gigabyte ternary version for laptops. Both were genuinely usable. Neither was genuinely competitive. The original ternary build held onto about 95% of its full-precision counterpart's benchmark score, impressive for a compression trick, still a real gap for anyone doing serious work.
Qwen3.8 Teases a 2.4 Trillion Parameter Open Model as Alibaba Chases Kimi K3 Alibaba's Qwen team teased Qwen3.8, a 2.4 trillion parameter model heading toward an open-weight release, with a live preview already on chat.qwen.ai. The announcement lands just as Moonshot AI's Kimi K3 forces Chinese AI labs into a race over who actually ships usable open weights first. - Alibaba Qwen3.8 open source model - trillion parameter language model release
Bonsai 2 closes most of that gap. Three percentage points sounds small until you remember what sits on the other side of it: the difference between a model that stumbles on harder reasoning and one that mostly doesn't. PrismML didn't get there with a new compression method. It got there by starting from a better base model, Qwen3.8 instead of Qwen3.6, and tuning the training recipe around the ternary constraint. The underlying trick is the same one Microsoft's BitNet researchers have pushed since 2024: constrain weights to -1, 0, or 1, and you strip out nearly every multiplication in the network, replacing it with plain addition. Multiplication is expensive in silicon. Addition is cheap.
What it means beyond one model's benchmark score #
That's the part that should worry chipmakers more than any single leaderboard number. For two years, the working assumption across AI has been that better models require bigger, pricier accelerators, the thesis behind hundreds of billions of dollars in Nvidia GPU orders and new data center buildouts. A 27-billion-parameter model that fits in under 6 gigabytes and runs at real speed on a five-year-old gaming GPU, or a phone, chips away at that assumption for a specific and growing category of work: running models people already have, not training the next frontier one.
It doesn't undercut the assumption everywhere. PrismML trained Bonsai 2 on Google's v5 TPUs, and training still eats enormous compute no matter how small the finished weights end up. The ternary trick shrinks what you need to run a model, not what you need to build one.
Frankly, that split is the whole story here. The AI industry's hardware demand for training frontier models isn't going anywhere. But the assumption that using a capable model requires a data center gets harder to defend with each release like this one.
Also read: Geoffrey Hinton Tells Senators Congress Has About a Year to Rein In AI • Anthropic Turns Claude Code Into a Multi-Agent Team With New Projects Beta • CoreWeave Plans Another $3.5 Billion In Convertible Notes To Fund AI Growth
This article is posted in AI News, check it out for more related stories.
Alibaba's Qwen3.8 Arrives With 2.4 Trillion Parameters After Kimi K3 Selloff Alibaba's Qwen team announced Qwen3.8, a claimed 2.4 trillion parameter model headed for open-weight release, just two days after Moonshot's Kimi K3 triggered a $3.3 trillion chip stock selloff. No independent benchmarks or open weights have landed yet, so the claim remains unverified. - Alibaba Qwen3.8 2.4 trillion parameters - open weight AI model release
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.