Moondream 3 對比 LFM2.5-VL 3B：一個指向答案，一個只是回答

The most useful difference between M​oondream 3 Preview and LFM2.5-VL-3B is not in their benchmark tables. It is that M​oondream 3 Preview reasons about an image by pointing into it — its reasoning traces are grounded, referencing specific regions — while LFM2.5-VL-3B is explicitly a non-reasoning model that answers directly, a choice Liquid AI made to keep latency low for on-device work. Same image in, two entirely different things come out, and which one you want depends on whether a downstream system has to trust the answer or merely consume it.

Everything else follows from that fork. M​oondream 3 Preview is a sparse mixture-of-experts model at 9B total parameters with 2B active, released on 18 September 2025 by M​oondream, which was acquired by Prime Intellect in October 2026. LFM2.5-VL-3B is a dense 3.1B model — a 2.6B language backbone with a 400M SigLIP2 NaFlex vision encoder — released by Liquid AI on 12 August 2026 as a direct successor to LFM2-VL-3B. One is a specialist that was built to detect and point and can also answer questions. The other is a generalist that was built to answer and recently learned to point.

你必須先注意到的指標碰撞

把兩個接地數值並排放置，比較起來幾乎不相上下。但事實並非如此，因為它們不是同一種量測。

M​oondream's launch chart lists M​oondream 3 Preview at 91.1 on RefCOCO. Liquid AI reports LFM2.5-VL-3B at RefCOCO precision@1 of 87.9. RefCOCO accuracy and RefCOCO precision@1 are different statistics computed over the same dataset, and a 3.2-point gap between them tells you nothing about which model localises better. Anyone stacking those two figures in a bar chart is comparing a number to a differently-defined number.

• M​oondream 3 Preview grounding — RefCOCO 91.1, on a chart where frontier baselines (GPT-5, Gemini 2.5-Flash, C​laude 4 S​onnet) were queried through a template because they have no native grounding, and where some preview scores came from a 100-question sample rather than the full set. Vendor-reported and, as M​oondream says itself, not a fair fight.

• LFM2.5-VL-3B 定位 — RefCOCO precision@1 87.9，較前一代 LFM2-VL-3B 的 57.1 提升，Liquid 將此歸因於擴大合成定位資料的規模。廠商自行回報，在相同測試框架下與自家前代相比，這是更具意義的比較軸線。

• Detection and pointing — M​oondream 3 Preview: native detect and point skills, structured coordinate output, part of the architecture. LFM2.5-VL-3B: grounding added by training, returned as coordinates the same way, with function calling also new to the VL line (ToolSandbox 26.4 to 59.5, BFCL v4 20.5 to 32.5).

The honest read: Liquid AI has a clean before-and-after on identical harnesses and M​oondream has a cross-model chart with template caveats. Neither gives you a head-to-head. If grounding accuracy is the deciding factor in your evaluation, your own labelled set is the only instrument that will settle it.

有憑有據的推理讓你付出什麼代價

M​oondream's query skill runs in reasoning mode by default. The model works through the question and its trace references the regions of the image it is using, which you can inspect in the playground by hovering over the text. For anything where a human has to audit the answer — a defect call on a production line, a security incident, a medical triage flag — that property is worth a lot. It is the difference between a model that says "no" and a model that says "no, here." p>The cost is tokens. Reasoning mode generates a trace before the answer, and M​oondream's own documentation offers the escape hatch: pass reasoning=False for simple questions, because the company knows the mode is not free. That is a real architectural admission — a well-designed model would not need a flag to switch the thought off.

LFM2.5-VL-3B 沒有對等的同類模型。它被設計為非推理模型，而 Liquid AI 也直言這項取捨：直接作答可為即時與裝置端使用維持低延遲。沒有可供檢視的軌跡，也沒有可停用的模式。在結帳櫃檯掃描條碼，或用手機翻譯菜單時，這正是正確的工程決策，而缺少軌跡並非缺陷。

So the question is not which is better. It is whether anything downstream of your model reads the reasoning. If a human does, M​oondream's grounded traces are the feature you are buying. If only a program does, you are paying for tokens nobody will read.

3.1B 真正勝出的地方

LFM2.5-VL-3B 的體積宣稱並非行銷話術。它在 Apple M5 Max 上只需不到 3.3 GB 的記憶體就能運行，速度達每秒 228 個 token；在 AMD Ryzen AI Max+ 395 上則是每秒 116 個 token，而 GGUF、ONNX 與 MLX 匯出格式也與原生 checkpoint 一併發布。這是一套能放進筆電的模型，Liquid AI 還將它打包成可在瀏覽器分頁中運行的 WebGPU Hugging Face Space。

M​oondream 3 Preview has a smaller active parameter count — 2B against 3.1B — but the comparison is not symmetric, because a MoE checkpoint has to be fully resident even though only a fraction of it fires. The 9B of weights need somewhere to live. M​oondream's own answer to that is Photon, its local inference engine, which runs the model on N​VIDIA GPUs from Ampere onward and on Apple Silicon Macs. Different constraint: LFM2.5-VL-3B is small enough to fit anywhere, M​oondream 3 Preview is fast in the compute sense but large in the memory sense.

• Context — both 32K tokens. Genuine convergence, and worth noting given the two-year jump from the 2K windows of the previous M​oondream generation.

• Languages — LFM2.5-VL-3B lists sixteen: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish, and doubled its vocabulary to 128,000 to support non-Latin scripts. M​oondream 3 Preview's card leads on visual capability rather than a language list; treat multilingual VQA as something to test rather than something promised.

• Document reading — LFM2.5-VL-3B: DocVQA 91.1, ChartQA 81.3, OCRBench v1 84.2, OCRBench v2 English 47.5, MMMU 48.4. M​oondream 3 Preview: DocVQA 88.3, ChartQA 86.6, POPE 89.0. Different harnesses, vendor-reported on both sides.

• LFM2.5-VL-3B 明白告訴你哪些事不該做——Liquid AI 的說明卡直言，不建議將它用於長上下文或推理密集的任務，例如視覺化網頁設計或關於藍圖的技術問題。這句話比大多數基準測試表格都還實用。

部署，以及路由器適合的位置

Neither of these models is served through OrcaRouter, and it is worth saying so directly rather than dancing around it: both are open-weight checkpoints you run yourself, and M​oondream 3 Preview's hosted endpoint — M​oondream Cloud — is being shut down on 13 November 2026 following the Prime Intellect acquisition, with unspent credits refunded.

這改變的是決策的形態。一個以自架 VLM 為標準的團隊，通常最後會再搭配一個託管的第二模型，來處理地端模型應付不來的情況，而到了那個時候，問題就出在整合工作上：兩套 SDK、兩套驗證機制、兩套重試策略，還有供應商的中斷會變成你的中斷。

This is the case OrcaRouter's routing DSL is built for. You compose several models into a single call — the 3.1B local model answers what it can, a hosted model takes the residue — behind one key for more than 200 models, with provider list price passed through at zero markup and automatic failover when an upstream degrades. Since neither M​oondream 3 Preview nor LFM2.5-VL-3B is in that catalogue, the routing layer sits above them rather than replacing them: it is how you keep the cheap local path cheap without making a vendor outage your outage.

授權分叉，比基準更為犀利

兩種不同的授權哲學，而這個差異對產品決策的影響，遠比3分的分數差距來得重要。

M​oondream 3 Preview ships under the Business Source License 1.1 with an Additional Use Grant (No Third-Party Service). Internal and production use, personal projects, research, benchmarks, fine-tunes, merges, quantisations, weight deltas, and use inside a product that does not substantially overlap M​oondream's paid offerings are all permitted. Selling a hosted API that exposes similar vision capabilities, managed "M​oondream-as-a-service", or embedding the weights in a paid SDK that delivers comparable capabilities requires a separate agreement. M​oondream 3.1 moved to the M​oondream Model License 1.0, which states the same rule more cleanly: component of your product is allowed, access to M​oondream itself as the product is not. Edge devices, customer-controlled VPCs, and running the model on leased cloud or neocloud GPUs are all explicitly on the allowed side.

LFM2.5-VL-3B is under Liquid AI's own LFM Open License v1.0. Read it before you ship, as you would any bespoke licence, but the families are different in kind: M​oondream's restriction targets one specific business — reselling the model — while Liquid's is a broad open licence with its own terms.

If you are a domain-specific builder, both are fine and the licence question is a formality. If your business is anywhere near "we host vision models for other people," M​oondream's licence is the constraint you read first and LFM's is the one you probably clear.

該選哪一個

Choose M​oondream 3 Preview if the answer has to be inspectable. If a human reviews the output, if the trace needs to show which part of the image drove the decision, if detection and pointing are the primary product rather than a feature — that grounded reasoning is the whole reason to pay 9B of memory for a 2B-active model. Budget for Photon or a GPU, and plan around the hosted endpoint closing.

選擇LFM2.5-VL-3B，如果模型必須駐留在裝置上。如果你要部署到手機、筆電、工廠現場機箱或瀏覽器分頁，而且工作負載是單輪、高吞吐量且低延遲——讀取螢幕、轉錄文件、偵測串流中的物件——那麼一個不假思索就能回答的稠密 3.1B 模型正是你想要的，而它在 M5 Max 上每秒 228 個 token 的數字，就是決定關鍵。

Pick both if you have the budget for it, and route between them: LFM2.5-VL-3B on every frame because it is nearly free to run, M​oondream 3 Preview on the frames where a grounded answer is worth waiting for.

數字無法告訴你的是，這一切在實際運作中是否出了問題：針對此類模型有記錄的 AI 事件與代理故障，都收錄在the OrcaRouter AI Incident Archive，其中每筆條目都附有自己的來源、嚴重程度、可信度等級與爭議標記。