Bonsai 2 27B on Mac: small download, larger footprint

Reference Mac: Apple M1 Max · 32-core GPU · 64 GB. Decode is engine-reported tokens per second; other Macs are not measured.

Compare profiles in the model catalog

Measured to 65,964 tokens (39.7 GB peak, M1 Max 64 GB). The 81,920-token window is estimated at 53.0 GB.

The catalog dated 2026-10-09 records an 8.6 GB download for the MLX 2-bit ternary profile. The reference Apple M1 Max has a 32-core GPU.

Weights are only part of a running model’s memory use. Attention state grows with the chat, and macOS and your coding apps need room too. Subtracting download size from peak memory would not produce a measured cache size.

Choose something else when you need a larger admitted window or the estimated window total leaves too little room for your work. The Qwen3.8 article compares two larger-window profiles; the requirements page explains how to budget memory. No fit on a smaller Mac is established by this result.

Prism ML identifies Bonsai 2 as derived from Qwen3.8 27B. The earlier Ternary Bonsai 27B is derived from Qwen3.6 27B. The version number marks a different base model, so an observation from the earlier pack should not be carried over to this one.

The earlier pack is an 8.5 GB download. Measured to 65,988 tokens (37.5 GB peak, M1 Max 64 GB). The 98,304-token window is estimated at 54.0 GB.

Bonsai 2 instead has the 8.6 GB download, 81,920-token window and 53.0 GB estimate shown here. These catalog entries share the 2026-10-09 date and reference Mac; they do not establish a same-task quality comparison.

The publisher describes ternary language weights with values of minus one, zero and plus one. Its model card distinguishes the ideal weight representation from the MLX pack, which also includes a vision tower. Those are descriptions of storage and representation, not measurements of Bridge1’s total memory use.

Prism ML’s retained-intelligence and compression claims belong to its own evaluations. Bridge1’s 53.0 GB estimate describes memory at the admitted window on the reference 64 GB M1 Max. It does not validate the publisher’s retained-intelligence claim or predict coding success in your project.

Depth decode is not measured for Bonsai 2. The short-decode cell has a recorded value with conditions not recorded, so it cannot establish long-chat latency. Choose a bounded task with a test you can check before committing to a longer session.

The official Bonsai 2 MLX page declares Apache-2.0, as does the earlier Bonsai MLX page. The linked pages were accessed 2026-10-19.

The MLX 2-bit ternary profile’s name in Bridge1 should read Ternary Bonsai 2 27B. The word “2” matters: the earlier Ternary Bonsai model is a separate choice.

Read the admitted window before starting a coding chat. Keep the download size separate from the window estimate: the chart above labels the memory evidence for this profile at the catalog window.

On macOS, bridge1 launch starts Claude Code, Codex, OpenCode or Pi with Bridge1, subject to the installed client version, model availability and loading settings; the manual steps still work. With Bonsai 2 selected, open Gateway and copy OpenCode’s complete provider entry. The connection values below come from the OpenCode guide. Replace the placeholder key and preserve any generated attachment, modalities and context fields.

Route1 follows the selected model. If you choose a plain model id instead, use the one Gateway supplies for Bonsai 2. Follow the guide for your client below; a successful connection alone does not establish a long coding-session result.

// Add this "bridge1" entry to the "provider" object in ~/.config/opencode/opencode.jsonc,
// and set "model" to "bridge1/route1"
"bridge1": {
  "npm": "@ai-sdk/openai-compatible",
  "name": "Bridge1",
  "options": {
    "baseURL": "http://127.0.0.1:54321/v1",
    "apiKey": "YOUR-GATEWAY-KEY"
  },
  "models": {
    "route1": {
      "name": "route1",
      "attachment": true,
      "modalities": {
        "input": [
          "text",
          "image"
        ],
        "output": [
          "text"
        ]
      },
      "limit": {
        "context": 262144,
        "output": 32768
      }
    },
    "mtp-qwen3.6-35b-a3b-gguf": {
      "name": "mtp-qwen3.6-35b-a3b-gguf",
      "attachment": true,
      "modalities": {
        "input": [
          "text",
          "image"
        ],
        "output": [
          "text"
        ]
      },
      "limit": {
        "context": 262144,
        "output": 32768
      }
    }
  }
}

Check client support and known limits