InclusionAI releases Ling 3.1 Flash with 560B parameters, 25B active and a 1M-token context

InclusionAI has released Ling-3.1-flash, an open-weight reasoning model with approximately 560 billion total parameters, 25 billion active parameters per token and an advertised one-million-token context. Developed by the Ant Group-associated AI initiative, the model combines sparse expert routing with hybrid attention to target coding agents, long-document analysis and tool-intensive enterprise workflows. Its weights are available under the MIT license, although running a checkpoint of this size remains a substantial infrastructure undertaking.

The release is notable for its engineering choices rather than parameter count alone. Ling-3.1-flash activates only 4.46% of its weights per token, combines 63 KDA attention layers with nine gated Multi-head Latent Attention layers and offers deployment recipes for SGLang and vLLM. Those characteristics could improve the economics of complex agentic work, but published specifications do not establish real-world superiority over competing models. In particular, the advertised one-million-token context must be distinguished from the 262,144-position setting in the public configuration.

··········

LING-3.1-FLASH: SPECIFICATIONS, OPEN WEIGHTS AND THE CONTEXT LIMIT.

Ling-3.1-flash is a hybrid-linear Mixture-of-Experts (MoE) model designed for reasoning and text generation. InclusionAI distributes the checkpoint through Hugging Face with configuration files, a chat template and inference instructions, allowing developers to deploy the model on their own infrastructure instead of relying exclusively on a hosted API.

........

Specification	Ling-3.1-flash
Developer	InclusionAI, associated with Ant Group
Release	October 2026
Architecture	Hybrid-linear Mixture-of-Experts
Total / active parameters	560B / approximately 25B
Active parameter ratio	4.46%
Advertised context	1 million tokens
Public maximum-position configuration	262,144
Transformer layers	72
KDA / gated MLA layers	63 / 9
Routed / shared experts	512 / 1
Routed experts selected per token	8
Dense layers	4
Attention heads / hidden dimension	64 / 4,096
Vocabulary size	157,184
Weight distribution	FP8 safetensors
Model license	MIT
Serving frameworks	SGLang and vLLM

........

The model's headline context figure describes an advertised capability, whereas the maximum-position setting describes the published checkpoint configuration. A provider may expose additional context through particular serving settings or extensions, but developers should not assume that every installation accepts one million tokens, or that accuracy remains constant across the entire window. Actual usable context, retrieval quality and throughput need to be tested separately.

··········

WHY 560B TOTAL PARAMETERS AND 25B ACTIVE PARAMETERS ARE DIFFERENT METRICS.

In a dense transformer, essentially all model weights participate in each token's computation. Ling-3.1-flash instead uses a router to select a subset of expert networks, activating approximately 25B ÷ 560B = 4.46% of its total parameters for a typical token. This sparse design allows a large pool of learned representations without performing a full 560B-parameter computation at every decoding step.

The architecture contains 512 routed experts, of which eight are selected for each token, plus a shared expert. Selecting eight out of 512 corresponds to 1.5625% of routed experts, a figure that should not be confused with the 4.46% active-parameter ratio because embeddings, attention and dense components also contribute to the model's computation.

The main trade-off is that lower active computation does not eliminate the memory footprint of the complete model. Expert weights must remain accessible, and routing tokens across GPUs can introduce communication overhead and uneven accelerator utilization. Consequently, comparing Ling's 25B active parameters with a dense 25B checkpoint would underestimate its storage and serving requirements; the relevant economic measurements are hardware occupancy, throughput, latency and cost per successful task.

··········

HYBRID ATTENTION IS CENTRAL TO LING'S LONG-CONTEXT DESIGN.

Ling-3.1-flash distributes its 72 transformer layers between 63 KDA layers and nine gated Multi-head Latent Attention (MLA) layers, producing a seven-to-one ratio. KDA contributes an approach intended to reduce the cost of maintaining information across long sequences, while MLA uses compressed attention representations to address memory requirements associated with conventional attention. Combining the two mechanisms seeks to preserve useful token relationships without incurring the full cost of a conventional attention design at every layer.

........

Attention component	Layers	Share
KDA	63	87.5%
Gated MLA	9	12.5%
Total	72	100%

........

This arrangement is relevant to agents that repeatedly append code, logs, search results and tool responses to their working context. Yet accepting a long sequence is not equivalent to understanding it reliably: information placed near the beginning, middle and end of a large prompt can be retrieved with different accuracy, while contradictory documents and irrelevant material may further reduce performance. For production deployments, a long-context test should include multi-document evidence retrieval, distant cross-references and reasoning under distracting information rather than merely confirming that an input fits.

For scale, a million tokens might correspond to approximately 1,333 pages at an illustrative 750 tokens per page; the actual page count depends heavily on language, formatting and code. The public 262,144-position configuration is roughly 26.2% of the advertised million-token figure, making verification of the deployed context limit a prerequisite for any application designed around the larger capacity.

··········

CODING AGENTS AND TOOL USE ARE THE MAIN APPLICATION TARGETS.

InclusionAI positions the model for software engineering, terminal tasks, search, office workflows and general-purpose agents. These applications require more than a plausible final answer: a coding agent may need to inspect a repository, modify files, run tests, interpret failures and repeat that sequence for hours. Its success depends on reliable tool arguments, state tracking and error recovery as well as the model's underlying reasoning.

The developer describes compatibility with Claude Code, Kilo Code, Qwen Code, Hermes Agent and OpenClaw, and has presented LingCode, a coding-agent harness based on the Pi framework. LingCode is intended to support lengthy workflows with many tool calls and uses context compaction to manage growing histories. Compaction can reduce repeated processing, although an overly aggressive summary may discard constraints or evidence needed later; performance therefore depends partly on the harness rather than the checkpoint alone.

InclusionAI cites Terminal-Bench 4.0, SWE-bench Pro, FrontierSWE, GDPVal-AA V2.1, AutomationBench, MCP-Atlas and SkillsBench among its evaluations, alongside professional-domain tests. These benchmarks measure different capabilities and may use different tool access, retry budgets and agent configurations. The developer's reported results are informative, but they should not be combined into a general claim that Ling outperforms every alternative on coding, finance or healthcare.

··········

BENCHMARK ACCURACY MUST BE WEIGHED AGAINST TASK COMPLETION COST.

Agent evaluations are sensitive to the number of allowed attempts, execution time, test access and context-management policies. A system that spends five times as much on each attempt may achieve a modestly higher completion rate without being economically preferable for routine enterprise work. To illustrate the distinction, consider two hypothetical coding agents:

........

Illustrative metric	Agent A	Agent B
Tasks attempted	100	100
Successful tasks	70	75
Completion rate	70%	75%
Average cost per attempt	$2	$10
Total expenditure	$200	$1,000
Cost per successful task	$2.86	$13.33

........

These are not measured Ling results. They show why cost per successful workflow can be more meaningful than completion percentage alone: Agent B solves five additional tasks but incurs substantially higher expenditure. A fair comparison of Ling with other models should use the same tasks, tools, retry limits and hardware-accounting assumptions, reporting completion rates alongside elapsed time and total inference and execution cost.

For business use, error severity also matters. A failed coding test can often be rerun, whereas a wrong financial reconciliation or an unauthorized tool action may create a materially different risk. Benchmark selection should reflect those consequences.

··········

FP8 WEIGHTS AND MULTI-GPU SERVING DEFINE THE DEPLOYMENT ECONOMICS.

The public checkpoint is distributed in FP8-related format. At approximately one byte per parameter, 560 billion parameters imply roughly 560 GB of raw weight storage; the repository itself is approximately 574 GB. The gap is consistent with additional files and representation overhead, while runtime memory can increase further because of activations, attention states and serving buffers.

........

Representation	Idealized raw storage for 560B parameters
FP32	2.24 TB
BF16 / FP16	1.12 TB
FP8	560 GB
4-bit	280 GB

........

These are theoretical conversions, not validated deployment footprints. Quantization may reduce storage but requires a compatible implementation and can affect model accuracy. In particular, the small active-parameter count does not make the complete checkpoint suitable for an ordinary single-GPU workstation.

InclusionAI's SGLang instructions describe deployments using eight NVIDIA H200 GPUs with tensor and expert parallelism, as well as a four-GB300 configuration. Tensor parallelism divides selected computations across devices, while expert parallelism distributes expert networks so routed tokens can reach the relevant weights. The model documentation also describes vLLM support, including expert parallelism, prefix caching and model-specific tool-call parsers.

Speculative decoding and prefix caching can improve throughput under favorable conditions. The former proposes tokens ahead for validation; the latter reuses computation when requests share an identical prompt prefix. Neither removes the underlying memory requirement, and gains depend on prompt length, batching, concurrency and how often prefixes are reused. Developers should benchmark time to first token, prefill throughput, decoding speed, concurrent capacity and end-to-end cost, not merely tokens per second in an isolated short prompt.

··········

WHAT THE MIT LICENSE ENABLES, AND WHAT OPEN WEIGHTS DO NOT PROVIDE.

Ling-3.1-flash's downloadable weights and MIT license allow developers to inspect the released artifacts and build private inference deployments, subject to the applicable terms. This is useful for organizations that need to control data flows, tune serving infrastructure or avoid sending confidential documents to a third-party inference API.

Open weights should nevertheless be distinguished from a fully reproducible development process. Public model parameters do not automatically include the complete training corpus, optimization history or every component of the developer's post-training pipeline. Likewise, the availability of the checkpoint does not establish that associated proprietary services or the LingCode harness are distributed under identical terms.

For enterprise deployments, licensing review should be paired with security testing, access controls and operational monitoring. Running the model locally changes where data is processed, but does not itself prevent prompt injection, unauthorized tool calls or errors in generated analyses.

··········

COMPARING LING-3.1-FLASH WITH META MUSE SPARK OPEN.

Meta Muse Spark Open has been described with 816B total parameters, 42B active parameters and an advertised one-million-token context. Comparing those published figures illustrates different architectural scale choices, but cannot establish which model is more accurate or cheaper to operate.

........

Specification	Ling-3.1-flash	Muse Spark Open
Total parameters	560B	816B
Active parameters	25B	42B
Active parameter ratio	4.46%	5.15%
Advertised context	1M tokens	1M tokens
Broad architecture	Hybrid-linear MoE	MoE

........

Ling has approximately 31.4% fewer total parameters and 40.5% fewer active parameters than these Muse figures. That difference may influence infrastructure requirements, but attention design, precision, expert placement, software optimization and output quality also matter. A credible comparison would use the same benchmark tasks, hardware, input lengths and cost-accounting method, especially for long-context and agentic workloads.

··········

FINANCIAL ANALYSIS AND PROFESSIONAL WORK REQUIRE DOMAIN-SPECIFIC VALIDATION.

Long-context models can be useful for comparing contracts, examining multiple financial reports and extracting evidence from large document sets. A financial controller might ask an agent to reconcile revenue across quarterly statements, identify changes in working-capital assumptions or compare contractual payment terms. Such tasks combine document retrieval with arithmetic and interpretation, so a plausible narrative is not enough.

For example, an agent must distinguish a year-to-date revenue figure from revenue earned in a single quarter, preserve the reporting currency and separate source figures from its own calculations. A mistake in any of those steps can invalidate the conclusion even if the model correctly retrieves the relevant paragraph. Evaluation should therefore measure numerical accuracy, evidence traceability, handling of contradictory documents and the rate of unsupported assertions.

The same caution applies to healthcare and scientific tasks. Reported benchmark gains may demonstrate performance on selected questions, but they do not establish clinical reliability or general scientific competence. High-stakes uses require validation against the specific workflow and appropriate human oversight.

··········

HOW TO EVALUATE LING BEFORE PRODUCTION DEPLOYMENT.

An effective pilot should reproduce the organization's actual tasks rather than rely exclusively on general leaderboards. For coding, that means real repositories, executable tests and recovery from tool errors. For long documents, it means evidence retrieval across different positions and conflicting source material. For financial workflows, it means reconciliations with independently checked answers.

........

Evaluation dimension	Recommended measurement
Reasoning and coding	Correct task completion
Tool use	Valid actions and error recovery
Long context	Retrieval and cross-reference accuracy
Numerical work	Calculation and reconciliation accuracy
Reliability	Failure and retry rates
Latency	Prefill and decoding time
Scalability	Concurrent tasks per GPU
Economics	Cost per successful workflow
Security	Prompt-injection and permission tests

........

Testing should use the intended serving stack and realistic input lengths. A model that performs well on a short, single-user benchmark may encounter different constraints when several agents simultaneously process large documents or execute tools.

··········

LING-3.1-FLASH EXPANDS THE OPTIONS FOR OPEN-WEIGHT AGENTIC AI.

Ling-3.1-flash combines a 560B-parameter sparse expert network, approximately 25B active parameters, hybrid KDA/MLA attention and publicly downloadable MIT-licensed weights. Its architecture is relevant to long-running agents and document-heavy workflows, while SGLang and vLLM support give developers options for deploying the large checkpoint across multiple accelerators.

The two most important qualifications are operational rather than promotional: the full model remains expensive to host despite sparse activation, and the advertised million-token context should be verified against the public 262,144-position configuration and measured long-context accuracy. InclusionAI's benchmark claims also need independent, workload-matched testing.

Ling-3.1-flash becomes commercially interesting when its architectural efficiency translates into reliable task completion at a competitive cost—not simply because it combines a large parameter count with a long advertised context window.

··········

FOLLOW US FOR MORE.

DATA STUDIOS

datastudios.org