
LingSante-3.1-air-VL vs MiniCPM-V 4.7: One Shows Its Work
- OrcaMỚIOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 trên 1 triệu token · 187 tok/s
- openaiMỚIOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 trên 1 triệu token · 87 tok/s
- openaiMỚIOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 trên 1 triệu token · 151 tok/s
- tencentMỚITencent: Hy4 Flash2026-11-16$0.02 / $0.07 trên 1 triệu token · 186 tok/s
- AlibabaMỚIQwen: Qwen4 Max2026-11-16$1.50 / $4.50 trên 1 triệu token · 85 tok/s
- GoogleMỚIGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 trên 1 triệu token · 705 tok/s
- AnthropicMỚIAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 trên 1 triệu token · 239 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 trên 1 triệu token · 96 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 trên 1 triệu token · 137 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 trên 1 triệu token · 85 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 trên 1 triệu token · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 trên 1 triệu token · 102 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 trên 1 triệu token · 70 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 trên 1 triệu token · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 trên 1 triệu token · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 trên 1 triệu token · 38 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Trí tuệ
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 trên 1 triệu token · 231 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Trí tuệ
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 trên 1 triệu token · 101 tok/s
Open both repositories on the same afternoon and the difference is not in the weights. LingSante-3.1-air-VL from inclusionAI publishes a twenty-six-row benchmark table with seven named comparators, a serving recipe for two inference engines, and a paragraph explaining which evaluation subsets were filtered. MiniCPM-V 4.7 from OpenBMB publishes five chart images, no per-benchmark numbers, and a single comparative sentence: that its 35B model reaches a higher overall score than Qwen3.6 35B-A3B in less inference time. One of these releases lets you check a claim. The other asks you to look at a picture of one.
That asymmetry is the whole comparison, and it is more useful than a spec sheet. The two models are not rivals in the ordinary sense: one is a 124-billion-parameter medical specialist that needs four datacentre GPUs, the other is a 35-billion-parameter generalist designed to run on a phone. They overlap on exactly one thing a buyer cares about, which is whether the vendor has given them enough evidence to commit. On that axis the medical model wins before a single score is read, and the rest of this piece is about what each one can actually prove.
Two MoEs, two very different bills
Both are mixtures of experts, and both lead with an active-parameter number that makes them sound smaller than they are. Read the storage line instead.
• Active parameters — LingSante-3.1-air-VL activates 5.5B of 124B per token; MiniCPM-V 4.7 35B-A3B activates roughly 3B of 35.2B. On compute per token the phone model is leaner.
• Stored weights — 249.7 GB of bfloat16 across fifty safetensors shards versus 70.4 GB across sixteen. On bytes resident in memory the gap is three and a half times, not the eleven times the active counts suggest.
• Serving floor — four 141 GB-class GPUs at tensor parallelism 4, or eight 80 GB H100/H800 at parallelism 8, for the medical model. OpenBMB's headline target is an iPhone 17 Pro Max at thirteen-plus tokens per second using about 3 GB, which is a vendor measurement nobody outside the lab has repeated.
• Context — 262,144 tokens on both. LingSante reaches it with a YaRN factor of 2.0 over a 131,072 base and documents the flag; MiniCPM declares 262,144 as its maximum position directly.
• Licence — MIT for the medical model, Apache-2.0 for MiniCPM. Both are permissive enough to deploy commercially, and neither carries a use rider beyond the research-only notice inclusionAI adds in prose.
The active-parameter trap is worth stating plainly because it decides deployments. A 124B model with 5.5B active still has to keep 124B of experts somewhere, and on a single device that somewhere is system RAM or a page file. LingSante does not pretend otherwise: its quickstart assumes tensor parallelism across four GPUs. MiniCPM's entire engineering story — lookahead expert scheduling, layer-wise quantisation, DSpark speculative sampling — exists to make the same arithmetic survivable on a handset. Both are honest about the constraint. Only one of them assumes you have a rack.
What MiniCPM-V 4.7 actually publishes

The OpenBMB card is well written and unusually specific about architecture. It names the vision tower, the 3:1 linear-to-full attention pattern, the mixed 4× and 16× visual token downsampling, and the exact devices the release targets: iOS, Android and HarmonyOS, with day-zero support for Rockchip, Houmo, Intel, AMD and Arm silicon. It lists SGLang, vLLM, llama.cpp and Ollama for inference, SWIFT and LLaMA-Factory for fine-tuning, and ships GGUF, BNB, AWQ, GPTQ and FP8 builds. As a deployment document it is better than most.
As an evaluation document it is thin in a specific way. The numbers live inside five linked images — an overall chart for the 35B vision-language tasks, a language chart, the 1B chart, a performance-efficiency frontier plot, and an on-device speed comparison. There is no table. That means the "higher overall score in less inference time than Qwen3.6 35B-A3B" claim cannot be checked per benchmark, and the composite behind it is OpenBMB's own weighting across groups OpenBMB selected. The company has shipped nine MiniCPM-V generations, several of them strong for their size, and this is not evidence of a weak model. It is evidence of an unaudited one.
Two facts outside the card do help. The repository has drawn more than 868,000 downloads in the past month and over 1,100 likes, which is adoption, not accuracy — people download models that are easy to download, and an Apache-2.0 checkpoint that runs in llama.cpp is easy. The 1B sibling is a separate, dense model with its own card and its own claims, and it should not be read as a compressed version of the 35B.
What the LingSante table proves
The medical card goes the other way. Its evaluation section is one image containing twenty-six rows and eight columns, but the rows are individually named and the comparators are individually named, which is the minimum required for a reader to disagree with a result. The comparator set is GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, MiniMax-M3.1-Flash, Kimi-K3.1-Pro, and three frontier systems at their highest reasoning effort: GPT-6.1 Sol, Claude Opus 5.5 and Gemini 3.8 Flash.
On the imaging rows the model posts CheXpert Plus 46.21, MIMIC-CXR 47.20, IU X-RAY 64.34, OmniMedVQA 93.64 and VQA-RAD 81.15, all vendor-reported and all first in that column of the table. It is second on PathVQA at 71.57, PMC-VQA at 65.82 and SLAKE at 86.24. Three of its firsts are on in-house sets — AFUMED-Derm 70.01, AFUMED-Food 72.80, AFUMED-Report 90.70 — which tell you what the model was trained toward and tell you nothing about how it compares to anything you can run yourself.
The text block is where the table turns. MedQA-USMLE 94.34 is last of eight. MedMCQA 81.54 is seventh. HealthBench Professional 49.59 is sixth, against 65.72 for the model's own 560B sibling. DiagnosisArena-MCQ 81.70 is fourth, its best text-only finish. A 5.5B active budget encodes what a radiograph shows; it does not carry the world knowledge that MedQA rewards, and the card's own numbers say so.
The document and agent rows are the surprise. OmniDocBench v1.5 at 91.20 edges Kimi-K3.1-Pro at 91.05 and Claude Opus 5.5 at 90.56. AndroidWorld at 86.84 sits above Kimi at 86.20 and well above the flash-tier models. Whatever this model is, it is not narrowly a chest-X-ray reader.
Medical rows nobody else runs
Here is the part a buyer should weigh before choosing either model on evidence quality. MiniCPM-V 4.7 does not run a single medical benchmark. Its published charts cover general vision-language, language and efficiency. If your task is a discharge summary, a dermatology triage queue or a radiology report draft, the phone model has no claim on it at all — not a weak claim, no claim. LingSante has claims on all three, and its own table shows it doing best where an image is attached.
The reverse is also true. LingSante does not report MMMU, DocVQA, ChartQA, OCRBench or any of the general multimodal suites that OpenBMB uses, so a team building a general document assistant has no basis for choosing it. OmniDocBench is the one document test it reports, and one row is not a general capability profile.
Neither model has an independent run. Neither appears on a neutral leaderboard with audited scores. The difference is that LingSante's vendor numbers are itemised and MiniCPM's are aggregated, and itemised numbers can be argued with.
Nơi mỗi thứ thuộc về
MiniCPM-V 4.7 is the right starting point when the constraint is the device: an offline capture app, a kiosk, a field tool with no reliable uplink, or a fine-tuning project with a small labelled set. Its breadth of quantisations and runtimes means a team can pick its stack rather than adopt one. The thirteen-tokens-per-second claim should be treated as a target to verify, not a specification.
LingSante-3.1-air-VL is the right starting point when the constraint is the task and the hardware already exists: a hospital or imaging vendor with 141 GB accelerators who needs a model that reads several images together, drafts Findings and Impression sections, and can call a bounding-box tool before it commits to a lesion location. It is a research checkpoint, not a medical device, and inclusionAI says so in the card. Deploying it as a clinical decision-maker would be a misuse of the artefact and of the licence's intent, even though MIT permits it.
The practical shape for most teams is a two-model pipeline: a perception stage that turns images into structured findings, and a reasoning stage that weighs those findings against guidelines. That second stage is the one worth routing. OrcaRouter composes several models into a single call through the routing DSL, so the medical vision checkpoint and a stronger text reasoner can come from different vendors behind one key, with automatic failover if a provider degrades. Neither LingSante-3.1-air-VL nor MiniCPM-V 4.7 is in the OrcaRouter catalogue today, and this article is not going to pretend otherwise; the point is that a 5.5B-active medical model is half a system even at its best, and the half it is missing is the half a router is good at.
The verifiable answer
If the question is which model is better, there is no answer in the public record, because nobody has run them on the same task. If the question is which release respects your time, the answer is the one with the itemised table. LingSante-3.1-air-VL tells you it scores 49.59 on HealthBench Professional, which is a bad number, and it tells you so in the same table where it claims a first place. MiniCPM-V 4.7 tells you it beats a model on an overall score and shows you a chart. For a team about to spend six figures on inference hardware, the willingness to publish the bad row is worth more than any single good one.

What would settle the rest is specific and short: a per-benchmark table from OpenBMB with actual figures, and one independent run of any LingSante row — HealthBench Professional first, because it is the row most likely to be quoted out of context. Until then the choice is made on deployment tier and disclosure, and on those two the models are not close.

So sánh trong bài viết này1
Phát hiện từ bài viết này · Benchmark: Artificial Analysis · cập nhật hằng ngày
