LFM2.5-24B-A2B vs Ternary Bonsai 2 27B: The Difference Is 2 Bits and a Cloud Connection

Here is the shortest honest answer. LFM2.5-24B-A2B is a text-only mixture-of-experts that Liquid AI built to sit inside an agent loop on a machine with no network, announced October 13, 2026 at 24.0B total and 2.45B active parameters. Ternary Bonsai 2 27B is a low-bit rewrite of Qwen3.8-27B that Prism ML released on September 17, 2026 under Apache 2.0, storing each weight as one of three symbols at 1.76 effective bits and shrinking a 53.81 GB FP16 model to a 5.93 GB download. They are both small enough to hold locally and both open-weight, and they are not trying to do the same job. Liquid's model was trained to call tools. Prism's model is a full-precision vision-language model wearing a much cheaper suit.

What makes the pairing worth a page is that each one tells you where its own evidence is thin, and the two weaknesses are opposites.

What each one is, reduced to the part that matters

LFM2.5-24B-A2B keeps the LFM2 architecture: 40 layers, 30 double-gated short-convolution blocks and 10 grouped-query attention layers, 64 experts per block with top-4 routing, 131,072 tokens of context, a 128,000-token vocabulary, text in and text out. Liquid's post-training pipeline ends with an agentic reinforcement-learning stage run inside OpenClaw, Hermes Agent, Pi and OpenCode. The vendor explicitly recommends pairing it with retrieval or web search instead of trusting it on knowledge questions — and on its own launch table the AA-Omniscience row sits at −24.90, which is that advice in numeric form.

Ternary Bonsai 2 27B is the same network as Qwen3.8-27B written down differently. Prism ML's whitepaper splits 27.36B parameters into 24.35B in the language backbone across 64 blocks, 2.54B in the embedding and LM head, and 0.47B in a 27-block vision tower that ships as a separate 4-bit mmproj file of roughly 0.63 GB and loads only when an image arrives. The base model's hybrid attention — 48 Gated DeltaNet linear-attention layers interleaved with 16 full-attention layers, hidden size 5,120, a 248,320-token vocabulary — is inherited unchanged. So is the 262,144-token context.

The line that separates them fastest is modality. Ternary Bonsai 2 27B accepts images. LFM2.5-24B-A2B does not, and no quantization trick changes that.

Quality, on the only table either of them has

Prism ML measured its build against the full-precision original across a 20-benchmark suite and reports it keeps 98.2% of the average — 83.9 against 85.4. That is the vendor's own harness, one month old, reproduced by nobody. It is also, deliberately or not, exactly the kind of claim that hides its own distribution, because the 1.8 points are not spread evenly.

• Math — 96.57 versus 97.06. Effectively level.

• Coding — 81.58 versus 82.17. Close, and the category the compression technique is being argued on.

• Instruction following — 82.66 versus 81.25. The compressed model is ahead, the one genuinely surprising line in the table.

• Knowledge and reasoning — 83.95 versus 86.66. A 2.7-point drop, and the single largest contributor to the missing 1.8.

• Agentic and tool calling — 77.57 versus 79.74, covering τ2-Bench at 80.22 and BFCL v3 at 74.92.

Now the problem with using that table here. Prism's numbers are its build against its own full-precision base. Liquid's numbers — AutomationBench 27.69, LiveClaw 48.09, ToolSandbox 80.71, BFCLv4 65.01, IFEval 96.75, IFBench 74.69 — are against a different set of models, on a different harness, with different generation settings. Not one benchmark appears in both tables. Anyone who puts BFCLv3 74.92 next to BFCLv4 65.01 and announces a winner has compared two different tests and called it a result.

Runtime is where the two diverge for real

This part is not a vendor claim, it is a compatibility fact you can hit in an afternoon.

Prism ML had to write kernels for the architecture, because the packed format and the mostly-linear attention are both unusual. Stock llama.cpp rejects both of Prism's packings as unknown types — and, more dangerously, loads an older ternary format without complaint and produces fluent nonsense, because no matching rotation has been applied to the activations. If you run Ternary Bonsai 2 27B on a binary that does not know about it, you do not get an error. You get confident, wrong output.

LFM2.5-24B-A2B has the opposite risk profile. Liquid lists day-one support across llama.cpp for GGUF, MLX, vLLM, SGLang and ONNX, plus FP8, ONNX and 4-bit MLX sibling repos — a conventional release. The open question there is performance rather than correctness: the vendor's figures of 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ 395 are unreproduced, and the headline "up to 1.7× faster than the next-best model" has no published baseline attached to it.

Where does 0% markup actually matter here?

Between these two models, the routing angle is genuinely awkward, and it is worth saying so rather than dressing it up. Ternary Bonsai 2 27B is a download: Prism ML distributes it on Hugging Face under Apache 2.0, and the point of the whole technique is that you run it yourself. LFM2.5-24B-A2B is also primarily a local artifact, with Liquid's Console endpoint at $0.10 per million input tokens — $0.05 cached — and $1.00 per million output as the hosted alternative. Neither one's core value proposition is a metered API call.

Where a routing layer earns its place is next to them, not instead of them. The base model Ternary Bonsai 2 27B is compressed from — qwen/qwen3.8-27b, published by A​libaba on August 13, 2026 — is served on our catalogue at $0.33 per million input and $2.40 per million output, which means the same task can run against the served full-precision model while you evaluate the 5.93 GB local build on the same prompts. That is a fair comparison being made cheaply, and it is the one Prism's own 20-benchmark table cannot make for you.

The licence, which is not a footnote on either side

Ternary Bonsai 2 27B carries Apache 2.0, with no revenue trigger and no field-of-use clause; the same is true of the Qwen3.8-27B weights it is derived from. LFM2.5-24B-A2B ships under Liquid's lfm1.0 licence, which is a different instrument with its own commercial terms. On a pure engineering read the two models are close enough that the choice often comes down to deployment shape — local executor versus local visual reasoner — but the licence is the one input that does not move with the benchmark cycle, and it is worth checking against your own revenue numbers before you build a product on either.

The thing each vendor left unmeasured is the thing to watch. Prism ML published a retention percentage but not the per-workload behaviour of a 1.76-bit model on your data, and a compression artifact that is 98.2% of average can still be the wrong tool for the one category you care about — the coding drop, small in aggregate, lands hardest somewhere specific. Liquid published an agentic table with no independent column and no long-context retrieval figures at all for a 131K window. Both gaps close with a benchmark release rather than a blog post, and whichever vendor ships one first will have moved the argument much further than the last 1.8 points did.

Track live model outages and degradations in our incident archive and see which routes have failed over cleanly.