Ling-3.1-flash: 560B Parameters, 25B Active, and What That Means for Your Hardware
Live timeline · last updated · 1 entry in the update log
Only 25B parameters work on each token, which sounds small. The weights still all have to live somewhere, and the recommended setup is a rack, not a desktop.
Ling-3.1-flash is a 560B-total, 25B-active model released about ten weeks after Ling-3.0-flash; it needs a multi-GPU server, and every benchmark number is the authors’ own.
Read from the inclusionAI/Ling-3.1-flash model card on Hugging Face on 2026-10-17; repository dates come from Hugging Face metadata. All benchmark figures are provider-reported and we did not reproduce them.
Timeline
-
The Ling-3.0-flash repository appears on Hugging Face. The 3.1 card later describes itself as building on this model, with improvements in architecture and in performance on real-world tasks.
-
The Ling-3.1-flash repository appears. The card introduces it as an upgraded native hybrid reasoning model for complex real-world tasks, with 560B total parameters, 25B active per token and a 1M-token context window, and ships day-0 serving recipes for SGLang and vLLM.
A release that starts with agents, not chat
The first paragraph of the Ling-3.1-flash card tells you what the authors want it to be used for. They call it an upgraded native hybrid reasoning model designed for complex, real-world tasks, and the examples they reach for are not conversation. They are general-purpose agents, search, office productivity and software engineering. The benchmarks the card names in its introduction follow the same pattern: GDPVal-AA V2.1 for office tasks, HealthBench Professional for healthcare, FrontierSWE for software engineering and Terminal-Bench 4.0 for terminal-based tool use. It also says the model improves in specialized domains such as finance and materials science, without numbers.
The lineage is short. The Ling-3.0-flash repository appeared on August 2. Ling-3.1-flash followed on October 13, roughly ten weeks later, and the card says the new model builds on the old one with improvements in architecture and in performance on real-world tasks. For anyone tracking the pace of open-weight releases, that gap is the headline: a 560B-parameter model revised in a little over two months.
Reading the specification sheet
The card’s overview table is where a model stops being a slogan. Ling-3.1-flash is a hybrid-linear mixture-of-experts model: 560B total parameters, 25B activated per token, 63 KDA layers and 9 gated MLA layers in a 7:1 ratio, 4 dense layers, 512 routed experts plus 1 shared expert, 8 experts activated per token, 64 attention heads, a hidden size of 4096 and a vocabulary of 157,184. The card does not expand the abbreviations KDA and MLA, and we are not going to guess at them. What the numbers show is arithmetic: 25B active out of 560B total is about 4.5 percent of the parameters touched for any one token, and 8 of the 512 routed experts, plus the shared one, do the work.
One discrepancy is worth flagging rather than hiding. Hugging Face’s own metadata on the page shows 572B parameters, a little above the card’s 560B. The difference is not explained anywhere we can see, so we use the card’s figure and note that the platform’s count is higher.
“25B active” is about speed, not memory
This is the sentence that gets lost in headlines about mixture-of-experts models. Compute per token tracks the number of active parameters, which is why a model with 25B active can be much cheaper to run per token than a dense model of 560B. But memory does not work that way. The router can send the next token to any expert, so all of them need to be resident in fast memory ready to answer. The weights therefore scale with the total, not the active count. At 16 bits per parameter, 560 billion parameters come to roughly 1.1 terabytes before the KV cache is counted. That is our arithmetic, not a figure from the card, and it explains the next section.
A long context window makes the memory story worse, not better. A 1M-token context is a large claim, and every token held in context adds KV-cache memory on top of the weights. The card does not give a memory figure for long-context use, so treat that as something to measure, not assume. A simple habit helps with any model of this size: the download is a floor, not a budget.
What the recommended setup tells you
The card’s recommended serving path is SGLang, and its day-0 image includes Ling-3.1-flash support, speculative decoding through MTP with the NEXTN algorithm, and Hopper decode optimizations. The server command is for 8 × H200 with tensor and expert parallelism both set to 8, or, on 4 × GB300, parallelism of 4. It adds that two optional environment flags give roughly 10 percent faster batch-1 decode on Hopper. A cookbook linked from the card lists verified recipes for H200, GB300, H800, H100 and B200, plus HiCache with Mooncake as a third cache tier, reasoning on and off, tool calling and measured accuracy and speed numbers. That is the profile of a model meant for data centers.
The vLLM route is an equally clear signal. The card asks you to build vLLM from a specific branch, then serve with tensor-parallel size 8, GPU memory utilization at 0.85, prefix caching, expert parallelism, a mamba cache mode, an auto tool-choice flag with a ling3 tool-call parser and a ling3 reasoning parser, and speculative decoding with three MTP tokens. Both routes use the same sampling recommendation: temperature 1.0, top-p 0.95, top-k 20, with thinking enabled for best performance.
Our reading is simple. This is not a model for a gaming GPU. If you want to estimate memory for models you do plan to run yourself, two habits help: assume runtime memory is larger than the file, and add the KV cache for the context length you actually use. For this particular model, the more practical question is whether a hosted endpoint serves it at a price that makes sense. The card points to OpenRouter and lists Novita as an inference provider, so hosted access exists, though pricing is not in the card.
What the authors claim it can do
The card says Ling-3.1-flash performs strongly on representative coding and agentic benchmarks: Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE, GDPVal-AA V2.1, Automationbench, MCP-Atlas and SkillsBench, among others. The evaluation block on the page lists scores for several: SWE-bench Pro 68.74, Deep SWE 59.7, Terminal-Bench 2.1 81.18, Humanity’s Last Exam 37.44, SWE-bench Multilingual (resolved) 88.5, Terminal-Bench 46.03 and SkillsBench v1.1 68.7. Several of these carry an asterisk on the page, and the page does not explain it in text we could read. These are provider-reported numbers. We did not reproduce them and they should not be read as a ranking against any other model.
The card also says the model gives a strong experience inside Claude Code, Kilo Code, Qwen Code, Hermes Agent, OpenClaw and other frameworks, a claim about compatibility rather than quality. And it introduces LingCode, a harness built on the open-source coding agent Pi and enhanced for long-horizon coding tasks that run for hours, involve hundreds of rounds of tool calls and require repeated context compaction. The authors say they are considering open-sourcing it, which means it is not released. That detail is easy to overlook and worth remembering: the strongest results the authors describe for long coding tasks use a harness you cannot yet download.
The paperwork
The card ends with something unusual for a model page: a public training-content summary, available as a PDF, and a documentation index. It describes the covered model versions, the training-content scope, the summary version and the update date, and says plainly that it concerns training-content disclosure and does not replace technical documentation, usage terms or the license. We did not read the PDF. The page also shows 1,904 downloads in the last month, a small number against the size of the model, which fits a release that is mostly served rather than downloaded.
Our take
Read the benchmark list as the authors’ statement of where they believe the model is strong, which is agentic coding and terminal work, and wait for independent runs before treating any single score as settled. Read the specification as a reminder that sparsity saves compute, not memory. The most useful number in the card for most readers is not a benchmark. It is the server command, because 8 × H200 tells you who this release is for. For anyone else, the question is not whether you can run it, but whether somebody serving it will give you access at a price that makes sense.
Keep reading on Front
Every post, newest first, is on the Front blog index.
- Reflection Beam 501B timeline: the dated timeline of the Reflection Beam 501B weights release.
- Meta Muse Spark Open timeline: the dated timeline of the Meta Muse Spark Open release.
- DeepSeek V4.1 Flash 1017 timeline: DeepSeek’s V4.1 Flash release, its Oct 17 update, prices and parameter counts.
- Humanizer 12B: which GGUF fits your memory: how much memory each GGUF file of the same model needs.
- Qwen Image 2.1 Turbo in eight steps: how Qwen Image 2.1 Turbo gets down to eight steps.
FAQ
How many parameters does Ling-3.1-flash have?
The model card says 560B total with 25B active per token, and a 1M-token context window. Hugging Face metadata on the page shows 572B.
Can I run Ling-3.1-flash on a single GPU?
Not on a consumer GPU. The card’s recommended setup is 8 × H200 or 4 × GB300, because all expert weights must be resident even though only 25B parameters are active per token.
What is Ling-3.1-flash good at?
The authors report strong results on agentic coding and terminal benchmarks such as SWE-bench Pro and Terminal-Bench. These are provider-reported and not independently verified here.
Update log
- — First published.