
LingSante-3.1-flash vs MiniMax M3.1 Flash: One Guesses, One Abstains
- OrcaNYOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 187 tok/s
- openaiNYOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 per 1M tokens · 87 tok/s
- openaiNYOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 per 1M tokens · 151 tok/s
- tencentNYTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 184 tok/s
- AlibabaNYQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 85 tok/s
- GoogleNYGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 705 tok/s
- AnthropicNYAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 239 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 96 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 137 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 85 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 123 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 102 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 70 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 per 1M tokens · 35 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 38 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligens
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 231 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligens
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 101 tok/s
On Artificial Analysis's Omniscience evaluation, MiniMax M3.1 Flash scores 16. That benchmark asks whether a model recognises the edge of its own knowledge, and a low score means it answers confidently where it should decline. LingSante-3.1-flash was trained, on its vendor's account, toward the opposite reflex: in the launch demonstration it computes heart rate and HRV per sleep stage and then leaves the field blank when the underlying signal cannot support a number. Put those two behaviours next to each other and most of this comparison is already decided. The parameter counts, the context windows and the per-token prices all matter, but they matter less than the fact that one of these models was built to answer and the other was built to stop.
Both are open-weight mixture-of-experts systems of roughly the same shape: a few hundred billion total parameters, about 23 to 25 billion active per token, a context window advertised at a million tokens. They share a release window of about five weeks. They are not competing for the same job.
What changed on 22 November
The weights for LingSante-3.1-flash appeared on Hugging Face at 04:30 UTC on 19 November, in the inclusionAI organisation, with no blog post and no announcement. By 22 November that had changed twice over. Ant Ling announced the model officially on its @AntLingAGI account, and shipped an MXFP4 quantised build alongside the original FP8 checkpoint. The quiet upload became a launch.
That timing matters for how you read the numbers below. Every medical score in this article comes from the vendor's own harness on the vendor's own card, and none of it has been reproduced. As of the announcement, the only independent figure on record is an Artificial Analysis MLCR of 23.3 percent, carried by the BenchLeader tracker, which also lists the model as having no benchmark index yet because it lacks the minimum number of independent results. The card says so itself in the section most readers skip: the model is intended for research and evaluation, it is not a medical device, and its outputs may be incorrect, incomplete or unsupported by the evidence they cite.

Specifikationsbladet, dimension för dimension
• Parameters — LingSante-3.1-flash 560B total, 25B active, against MiniMax M3.1 Flash 428B total, 23B active. Both are sparse; neither price is set by the total.
• Architecture — LingSante-3.1-flash is hybrid-linear: 63 Kimi Delta Attention layers and 9 gated MLA layers in a 7:1 ratio, 512 routed experts with 8 activated per token plus one shared expert. MiniMax M3.1 Flash uses sparse attention throughout and quantises experts and attention to 8-bit during training.
• Input — MiniMax M3.1 Flash takes text, image and video. LingSante-3.1-flash takes text.
• Context — both advertise 1,000,000 tokens. LingSante's shipped configuration is 262,144; the longer figure needs a serving recipe the card does not ship. MiniMax M3.1 Flash runs at the full million on first-party serving and drops to 256K on some hosts.
• Licence — LingSante-3.1-flash is MIT with no acceptable-use rider. MiniMax M3.1 Flash's weights were released after its API launch; check MiniMax's own terms before redistributing.
• Hardware — LingSante-3.1-flash needs roughly 578 GB of FP8 weights across eight H200s, or 291.72 GiB in MXFP4 across four H20-3e GPUs. MiniMax M3.1 Flash is the lighter serving problem on both counts.
• Price — MiniMax M3.1 Flash is $0.15 per million input tokens and $0.60 per million output, with cached input at $0.015. LingSante-3.1-flash has no published per-token rate card.
Where MiniMax M3.1 Flash is stronger
The gap is not close on agentic coding, and it is the largest single differentiator after modality. On DeepSWE v1.1, MiniMax reports 73.7 percent. On NL2Repo it reports 61.2. On OSWorld 2.0 it reports 61.7, on CyberGym 82 percent, and on BankerToolBench 76.2 — all vendor figures, all measured on MiniMax's own selection of harnesses, and all well above what a medical specialist is built to do.
It is also the faster model in practice. Artificial Analysis measures 135 output tokens per second, against a vendor claim of 170-plus median and 200-plus peak, and roughly three times the throughput of the M3 it replaces. Long agent loops spend most of their wall clock waiting on generation, so that difference compounds over a hundred-step trajectory in a way a single benchmark score does not show.

And it sees. A clinical pipeline that receives a faxed lab sheet, a photograph of a medication box or a scanned discharge summary can hand the file to MiniMax M3.1 Flash directly. LingSante-3.1-flash cannot accept it at all.
Where LingSante-3.1-flash is stronger
On the medical rows, the specialist wins by margins that are large enough to survive the vendor-reporting caveat. HealthBench Professional comes in at 65.72, ahead of GPT-6 Astra at 64.18 and Claude Opus 5.5 at 61.19 in the same table. MedQA-USMLE reaches 94.50. EHR-Complex is 74.40 and MedXpertQA-Text is 57.27. PhysicianBench lands at 57, second behind Gemini 4 Argon's 58. MedSCE, a safety benchmark, is 83.11.
The comparison that is actually like-for-like is narrower, and worth stating precisely because it is the one most write-ups get wrong. On Humanity's Last Exam without tools, LingSante-3.1-flash reports 44.07 and MiniMax M3.1 Flash reports 43.4 — effectively level, and both far below the 72.22 that LingSante reaches when tools are allowed. Its 94.10 on BrowseComp is likewise a with-context figure. Those headline numbers describe a model that is good at searching, not one that knows more.
So the honest reading of the general capability gap is: near parity on a hard no-tools reasoning set, a clear deficit on agentic software work, and a large advantage on medical professional behaviour that only one side measures.
The hallucination number that decides it
An Omniscience score of 16 is not a rounding error. Artificial Analysis uses that evaluation to separate a model that knows an answer from a model that produces a plausible one; a low score indicates a model that fills the gap rather than flagging it. In a customer-support bot that is an annoyance. In a system summarising a medication history it is the failure mode that matters most, and it is the failure mode a clinician cannot audit from the outside, because the wrong answer reads exactly like the right one.
LingSante-3.1-flash attacks that directly. Its training pipeline includes a behaviour matrix that scores responses against professional conduct criteria, with physician annotation agreement reported at 99.62 percent on that matrix, and the launch demonstration is built around refusal-shaped outputs: danger signs flagged, metrics left blank, recommendations revised as new examination data arrives. MedSCE at 83.11 and MedEthiAlign at 80.85 are the vendor's attempt to put a number on that behaviour.
Two caveats keep this from being a clean win. First, the abstention behaviour is demonstrated on curated examples, not measured across a distribution. Second, a medical-tuned 560B model can still be confidently wrong — its own card says so — and no independent party has tested whether it declines at the right moments. What can be said is that the design intent points in the right direction and the general-purpose alternative points away from it.
What a case costs on each side
MiniMax M3.1 Flash is the easy one to budget. Take a 30,000-token patient case — a long history, a set of labs, a guideline excerpt — and ask for 2,000 tokens of reasoning. Input costs $0.0045, output costs $0.0012, and the case lands at about $0.0057. Add prompt caching on the guideline section and the input side falls to roughly a tenth, which is what the $0.015 cached rate is for. At 10,000 cases a day that is a bill in the low hundreds of dollars, not thousands.
LingSante-3.1-flash cannot be priced that way, because Ant Ling has not published a per-token rate. Access at launch, on the vendor's own materials, runs through a Computrix token plan — the basic tier is quoted at US$20 a month — and the alternative is to serve the weights yourself. That second path is the interesting one. If you already own four H20-3e GPUs, the MXFP4 build fits in 291.72 GiB and the marginal cost of a case is electricity. If you do not, the same arithmetic that makes the model attractive on accuracy makes it expensive on capital: eight H200s is a commitment measured in six figures, and it buys you a text-only model.
The practical consequence is that these two do not sit on the same cost curve at all. One scales with volume at a known rate; the other has a fixed floor and near-zero marginal cost. Below a few thousand cases a month, the metered model is cheaper by a wide margin. Above the point where your GPU-hours are already paid for, the specialist becomes close to free per call.
MiniMax M3.1 Flash is on the OrcaRouter catalogue at exactly its list price — $0.15 in, $0.60 out, $0.015 cached — because we pass provider pricing through at 0 percent markup rather than adding a platform fee. That also means a MiniMax price change reaches your bill the same day it is announced, without a migration. LingSante-3.1-flash is not on our catalogue; it runs on the vendor's plan or on your own SGLang deployment, with the card recommending tensor and expert parallelism across eight H200s.
Vem ska köra vilken
If your job is intake — reading a scanned document, extracting a medication list from a photograph, triaging a queue of messages before a human sees them — MiniMax M3.1 Flash is the better tool, and the modality gap alone settles it. It is also the better tool for any agentic loop that has to call tools, recover from errors and keep going for dozens of steps, because that is what its benchmark selection measures and what its throughput supports.
If your job is the reasoning step — a differential, an evidence check, a summary that will be read by a clinician who expects the model to say when it does not know — LingSante-3.1-flash is aimed squarely at it. The accuracy claims are unreproduced, which means the responsible deployment is the one where a human reviews the output and the model's abstentions are logged and audited. That is a lower-volume, higher-stakes seat than intake, and it is the seat the model was trained for.
The two are complements more than substitutes, which is the unglamorous answer. A pipeline that puts M3.1 Flash in front of the document and something medical behind it is a reasonable architecture. What is not reasonable is choosing the generalist for the medical step because it is cheaper per token, and then discovering that the cheapness was measured on questions the model should have declined.
What to watch is the first genuinely independent medical evaluation of LingSante-3.1-flash. Until one exists, the vendor's table is the whole evidence base, and the safest assumption is that the abstention behaviour is real in the demonstrations and unquantified everywhere else.

Jämförda i den här artikeln1
Identifierat från den här artikeln · Benchmarks: Artificial Analysis · uppdateras dagligen
