Open Model WeightsINCLUSIONAI · FIELD VERIFIED
Primary sources last checked on 24 Oct 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
FRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-10-23
Ling 3.1 Flash is InclusionAI's first-party open-weight hybrid-linear mixture-of-experts language model for advanced reasoning, coding and tool-using agents. The developer reports 560 billion total parameters, 25 billion active per token and a configurable one-million-token context. The public Hugging Face repository declares MIT and includes model weights, custom code and serving guidance. Artificial Analysis separately identifies Ling 3.1 Flash as an open-weight model and evaluates its performance against comparable systems, providing an external curation signal without replacing checkpoint-specific verification. Open Model Weights source-verified the exact repository revision and license declaration before publication; it did not execute independent inference tests.
WHY IT MATTERS
The model combines a very large sparse expert pool with just 25B publisher-activated parameters per token, creating an interesting high-capacity inference tradeoff.
Independent Artificial Analysis open-weight evaluation provides a more defensible Leading signal than unreplicated publisher benchmark graphics.
The official SGLang serving notes and long-context architecture permit technically reproducible agent workload studies.
MODEL-SELECTION VIEW
The publisher describes a 560B total, 25B active hybrid-linear Mixture-of-Experts architecture, including 63 KDA layers and nine gated MLA layers. Routing uses a limited subset of the expert pool per token, which can reduce arithmetic versus dense activation while leaving total model state and checkpoint transfer costs substantial.
For comparison, report independently what the source model card says, what the configuration declares, and what weight files actually contain. Precision and sharding dominate practical GPU memory. A small active-parameter number does not let a multi-hundred-billion-parameter checkpoint fit on a single commodity graphics card.
InclusionAI advertises support for up to one million tokens and focuses on long-horizon coding agents, search and productivity workflows. The very large context can be useful for preserving prior tool results, but long context can also amplify irrelevant text, prompt injection and compaction defects.
Test retrieval precision near the beginning, middle and end of a long trace, not only a synthetic single fact. Monitor prefilling latency, cache allocation, truncation, tool state and the quality of the final answer. Use smaller production context windows until measured workloads justify the upper bound.
The first-party model card describes workflows with common coding agent frameworks and a LingCode harness based on Pi. Tool execution outcomes depend on schemas, argument formatting, user permissions, retries and the implementation of the external harness, not merely model weights.
For reproducibility, keep the exact agent runtime version, tool schema, sandbox permissions and benchmark tasks constant across candidates. Capture incorrect edits, test failures, unauthorized file access and recovery from malformed tool calls. A leaderboard win without a fair task budget does not establish safer autonomous behavior.
Artificial Analysis currently classifies Ling 3.1 Flash as open weights and reports 41 on its Intelligence Index in a large open-weight comparison. That supports a conditional editorial Leading placement after the canonical checkpoint passes source verification, rather than turning the score into an Open Model Weights measurement.
Vendor benchmark figures for terminal workflows, software engineering and productivity use different harnesses, temperatures and scoring. The independently evaluated service may still differ in quantization, batching and tool orchestration from a self-hosted implementation. Our model page should link both evidence types and avoid presenting the external score as a universal ranking.
The official repository declares MIT, publicly lists Safetensors shards and documents SGLang inference including distributed tensor/expert parallelism. Such evidence supports source-first availability and declared commercial rights, but deployment still requires a policy review of the exact code, data, artifact provenance and intended jurisdiction.
The published serving example uses eight H200 accelerators, illustrating the scale of practical inference in the author's setup, not a proven minimum requirement. Match custom architecture support and speculative decoding settings to the pinned runtime. If an accelerator stack lacks that implementation, a compatible model card is not proof of operational support.
The most useful comparison is to existing canonical Leading agent models in the same task class, for example GLM 5.3 or MiniMax M3. Standardize held-out coding changes, information retrieval, tool-call reliability and per-task cost, then keep raw logs and final patch acceptance criteria.
Begin with an isolated non-production trial and expand only when the model demonstrates repeatable quality under realistic concurrency. Score adversarial prompt injection, data leakage and tool mistakes separately from general reasoning ability. Treat all runtime benchmarks as unmeasured by Open Model Weights until a documented test actually runs.
IMPLEMENTATION & EVALUATION
Four-step source-pinned, reproducible evaluation plan tailored to Ling 3.1 Flash; not independently executed by Open Model Weights.
Capture 40-character revision, weight manifests, config, MIT metadata and custom implementation dependencies.
Immutable provenance manifest and licensing check.
A later upstream push can change evidence.
Reproduce a pinned SGLang multi-accelerator setting with fixed expert and tensor parallelism.
A deterministic generation and valid tool-call smoke test.
Advertised active parameters understate residency and cache.
Use held-out coding, retrieval and tool-use tasks with equal tools, retries and budgets.
Task-level quality, tool-call validity and injection-failure scorecards.
Benchmark harness differences invalidate broad comparisons.
Record actual peak VRAM, throughput, prefilling, p95 latency and failover with chosen context.
A published acceptance report and fallback trigger.
One-million-token configuration can be prohibitively expensive.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
InclusionAI's hybrid-linear MoE open-weight model targets reasoning, text-based agents, coding and long-context tasks with a claimed 560B total and 25B active parameters.
The first-party Hugging Face repository declares MIT and lists downloadable weights. A targeted canonical verifier still pins the precise revision and files.
The independently evaluated Ling 3.1 Flash model is described as text input and text output. Do not infer multimodal operation from another Ling family checkpoint.
Artificial Analysis independently lists Ling 3.1 Flash among high-performing open-weight models, reporting an Intelligence Index score of 41. The rank is editorial and does not override source-verification requirements.
The publisher demonstrates a distributed multi-accelerator setup. The 25B active-parameter number is not the whole model's storage requirement or a one-GPU guarantee.
One million tokens is the publisher's supported configuration target; deployers should measure long-context recall, cache memory, latency and failure rates for their own serving stack.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Ling 3.1 Flash as a 572.121B-parameter text to text, reasoning, coding and agentic tool use open-weight model in the Ling 3.1 family, with InclusionAI listed as the developer. The checked source repository provides 55 recognized weight artifacts, format evidence for Safetensors and a verified context window of 262,144 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares mit. Commercial use allowed under mit conditions. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
No base model is declared in the checked standard lineage metadata, so Open Model Weights does not infer one. Training recipe status: training-described-in-model-card; data disclosure status: not-disclosed-in-standard-metadata-or-obvious-model-card-section.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
Commercial use allowed under mit conditions
This section summarizes the checked repository evidence. It is not legal advice.
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
01 · WEIGHTS
model-00001-of-00054.safetensorsmodel-00002-of-00054.safetensorsmodel-00003-of-00054.safetensorsmodel-00004-of-00054.safetensorsmodel-00005-of-00054.safetensorsmodel-00006-of-00054.safetensorsmodel-00007-of-00054.safetensorsmodel-00008-of-00054.safetensorsmodel-00009-of-00054.safetensorsmodel-00010-of-00054.safetensorsmodel-00011-of-00054.safetensorsmodel-00012-of-00054.safetensorsmodel-00013-of-00054.safetensorsmodel-00014-of-00054.safetensorsmodel-00015-of-00054.safetensorsmodel-00016-of-00054.safetensorsmodel-00017-of-00054.safetensorsmodel-00018-of-00054.safetensorsmodel-00019-of-00054.safetensorsmodel-00020-of-00054.safetensorsmodel-00021-of-00054.safetensorsmodel-00022-of-00054.safetensorsmodel-00023-of-00054.safetensorsmodel-00024-of-00054.safetensorsmodel-00025-of-00054.safetensorsmodel-00026-of-00054.safetensorsmodel-00027-of-00054.safetensorsmodel-00028-of-00054.safetensorsmodel-00029-of-00054.safetensorsmodel-00030-of-00054.safetensorsmodel-00031-of-00054.safetensorsmodel-00032-of-00054.safetensorsmodel-00033-of-00054.safetensorsmodel-00034-of-00054.safetensorsmodel-00035-of-00054.safetensorsmodel-00036-of-00054.safetensorsmodel-00037-of-00054.safetensorsmodel-00038-of-00054.safetensorsmodel-00039-of-00054.safetensorsmodel-00040-of-00054.safetensorsmodel-00041-of-00054.safetensorsmodel-00042-of-00054.safetensorsmodel-00043-of-00054.safetensorsmodel-00044-of-00054.safetensorsmodel-00045-of-00054.safetensorsmodel-00046-of-00054.safetensorsmodel-00047-of-00054.safetensorsmodel-00048-of-00054.safetensorsmodel-00049-of-00054.safetensorsmodel-00050-of-00054.safetensorsmodel-00051-of-00054.safetensorsmodel-00052-of-00054.safetensorsmodel-00053-of-00054.safetensorsmodel-00054-of-00054.safetensorsmodel.safetensors.index.json02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
MODELS FOR AI AGENTS · ROLE EVIDENCE
Reasoning model. A candidate for a reasoning or coordination role. No end-to-end agent performance is inferred from its model category.
The curated leading tier is not an agent-readiness rating. Model-specific tool-call success, long-horizon task accuracy and compatible runtime configurations are not independently established here.
Source-linked license: mit (allowed commercial-use classification; see original terms).
Context: 262144 configured tokens.
BF16 weights: 1144.24 estimated decimal GB, not runtime VRAM.
Compare agent model candidates → · Agent evaluation method → · Inspect original model repository ↗
These records are separately sourced. They do not replace independent model-plus-runtime tests.
No model-specific agent capability evaluation is recorded for this checkpoint. Missing evidence is unknown, not a failed task.
Read evidence details and benchmarks → · Use the agent model selector →