
LingSante-3.1-air-VL vs Moondream 3: No Endpoint Either Way
- OrcaחדשOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 לכל 1M טוקנים · 187 tok/s
- openaiחדשOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 לכל 1M טוקנים · 87 tok/s
- openaiחדשOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 לכל 1M טוקנים · 151 tok/s
- tencentחדשTencent: Hy4 Flash2026-11-16$0.02 / $0.07 לכל 1M טוקנים · 186 tok/s
- AlibabaחדשQwen: Qwen4 Max2026-11-16$1.50 / $4.50 לכל 1M טוקנים · 85 tok/s
- GoogleחדשGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 לכל 1M טוקנים · 705 tok/s
- AnthropicחדשAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 לכל 1M טוקנים · 239 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 לכל 1M טוקנים · 96 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 לכל 1M טוקנים · 137 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 לכל 1M טוקנים · 85 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 לכל 1M טוקנים · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 לכל 1M טוקנים · 102 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 לכל 1M טוקנים · 70 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 לכל 1M טוקנים · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 לכל 1M טוקנים · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 לכל 1M טוקנים · 38 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447אינטליגנציה
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 לכל 1M טוקנים · 231 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929אינטליגנציה
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 לכל 1M טוקנים · 101 tok/s
On 13 November 2026, api.moondream.ai stopped answering. The notice that preceded it gave thirty days and promised refunds for unspent credits, and it applied to both Moondream 3 Preview and its successor Moondream 3.1. Ten days later, the comparison that matters between Moondream's small vision mixture-of-experts and inclusionAI's LingSante-3.1-air-VL is no longer "which one is more accurate." It is that neither model has a vendor endpoint you can call today. One lost its endpoint to an acquisition. The other never had one. Everything else follows from the size of the machine you are willing to own.
That is a stranger pairing than it first looks. Moondream 3 Preview shipped on 18 September 2025 as a 9-billion-parameter sparse model that activates 2 billion per token, and for most of its life the recommended way to use it was to buy inference. LingSante-3.1-air-VL shipped on 19 November 2026 as a 124-billion-parameter medical specialist that activates 5.5 billion, and its card links no API at all — only a four-GPU serving command. Ten days ago those were opposite business models. Today they are two answers to the same question: what does it cost to run this yourself?
What shut down on 13 November
Prime Intellect announced on 13 October 2026 that it had acquired Moondream, with founders Vik Korrapati and Jay Allen joining to lead inference. Moondream's own post followed on 14 October and confirmed the part that mattered to customers: Moondream Cloud, the hosted API that served both the 3 Preview and 3.1 checkpoints, would close on 13 November, with unused purchased credits refunded. The weights were not withdrawn. The licence did not change.
What went away is the layer that made Moondream easy. Its pitch to small teams was never that a 9B model beats a frontier system; it was that you could send an image and get a caption, a count or a pointing coordinate in tens of milliseconds without provisioning anything. Photon, Moondream's own inference engine, is the technical reason that pitch was credible — it compiles a model into a single GPU program with custom kernels and a megakernel compiler, then runs it locally on NVIDIA GPUs from Ampere onward and on Apple Silicon Macs behind the same API surface the cloud used. Prime's framing of the deal is explicit that this is what it bought: serving is where a model's usefulness gets decided, and the throughput numbers it cites — 34 requests per second for Moondream 3.1, more than 60 Moondream inferences per second on a single H100 — are now the product.
Those throughput figures are Prime's, and Prime is now the party selling the inference. Treat them as vendor-reported. What is not a claim is the shape change: a company whose product was a vision model bought an inference engine and put its authors in charge of inference. If you had a production call to Moondream's cloud, you have already migrated or you have already broken.
9B against 124B on the hardware bill
The architectural contrast is the useful part of this matchup, because it is the one thing both cards state clearly.
• Total parameters — Moondream 3 Preview stores 9B; LingSante-3.1-air-VL stores 124B. Thirteen and a half times the checkpoint.
• Active per token — 2B versus 5.5B. The medical model does roughly three times the per-token compute, which is the smaller of the two gaps.
• Expert layout — Moondream uses 24 layers, the first four dense, then MoE blocks with 64 experts and 8 active, hidden dimension 2,048. LingSante uses 42 layers, the first two dense, then 512 routed experts with 8 active, hidden size 2,560, plus a 27-layer vision tower.
• Resident memory — the practical number. Moondream's 9B fits on a single modern GPU or a large-memory Mac. LingSante's 249.7 GB of bfloat16 does not fit on one 141 GB accelerator: its own recipe uses tensor parallelism 4 across four of them, or parallelism 8 across eight 80 GB H100/H800s.
• Licence — the Moondream Model License 1.0, which is source-available rather than OSI-open, versus MIT with a research-only notice in prose.
The active-parameter ratio flatters the small model by a factor of nearly five and the memory ratio flatters the large one by three and a half. Neither number describes the deployment. The resident set does, and the resident set is why Moondream could be a laptop product and LingSante is a rack product.
The numbers each side publishes

Moondream 3 Preview's card is built around reference detection, counting and OCR, with a chart that the vendor annotates honestly: frontier baselines were queried through a template because they have no native grounding mode, and some preview scores came from a hundred-question samples. The repository has drawn about 37,500 downloads and 700 likes in fourteen months, which is real adoption for a model of its size. It has no medical rows. It was never aimed at them.
LingSante-3.1-air-VL publishes twenty-six rows, all medical or medical-adjacent, against seven named comparators. Its imaging firsts are CheXpert Plus 46.21, MIMIC-CXR 47.20, IU X-RAY 64.34, OmniMedVQA 93.64 and VQA-RAD 81.15. Its text-only block is weak: MedQA-USMLE 94.34 is last of eight, HealthBench Professional 49.59 is sixth, and its own 560B sibling scores 65.72 on that row. All vendor-reported, none independently reproduced.
Reading across the two tables produces exactly one honest conclusion: they share no benchmark, no modality emphasis and no intended deployment. Moondream is a perception utility for general images. LingSante is a clinical reading model with a document and agent tail. Putting their scores side by side would compare a captioning detector to a radiology report writer.
Photon is the actual product
There is a version of this comparison that treats Moondream 3 Preview as the loser because its cloud closed. That reading misses what Prime Intellect bought. Photon runs the same checkpoints locally with the same API contract, which means a team that was paying per image can move to a GPU it already owns and keep its integration. For counting, pointing and OCR over product photos or scans, a 9B model with a dedicated kernel compiler is a better fit than a 124B medical specialist will ever be, and the migration is a base-URL change rather than a rewrite.
The catch is that Photon is a runtime, not a service. Somebody has to own the GPU, the driver stack and the uptime. That is the trade Moondream's customers made on 13 November, whether they wanted to or not.
LingSante's card makes the opposite bet. It ships weights, config, chat template, a day-zero SGLang image and an inclusionAI vLLM fork, with a YaRN invocation that doubles a 131,072-token base to 262,144. It assumes you have the accelerators and it gives you the exact command. What it does not give you is a price, a rate limit, or a hosted path of any kind.
Where a router fits now
Both of these are now self-host propositions, so a routing layer is not a substitute for either one — it is what sits around them. The pattern that makes sense for a medical pipeline is a local perception stage and a remote reasoning stage: LingSante reads the images and emits structured findings, and a stronger text model weighs those findings against guidelines. A single key across 200-plus models, with automatic failover when a provider degrades, is a cheaper way to hold that second stage than signing a second contract, and the provider list price passes through at 0% markup, so a vendor price cut reaches the same endpoint the day it lands. Moondream 3 Preview is not in the OrcaRouter catalogue, and neither is LingSante-3.1-air-VL; the routing argument here is about the half of the system neither card provides.
Pick by the box you own

If you have one GPU and a stream of general images, Moondream 3 Preview plus Photon is the more sensible starting point, and its lost endpoint changes your billing model rather than your capability. Budget for the operational work you just inherited: the model no longer comes with a status page that is somebody else's problem.
If you have four 141 GB accelerators and clinical images, LingSante-3.1-air-VL is the only one of the two with a claim on your task, and its research-only notice should be read as a deployment constraint rather than boilerplate. Run it in evaluation, keep a human in the loop, and do not wire it to a diagnosis path.
If you have neither, the honest answer is that this matchup has no winner for you yet. The cheap path closed ten days ago; the expensive path was never open. What to watch is whether Prime Intellect restores a hosted Moondream endpoint under its own brand, and whether inclusionAI ever prices the medical model. Either event would turn this from a hardware decision back into a service decision.

