Open Model WeightsMINIMAX · 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
MiniMax M3.1 Flash is MiniMax's downloadable multimodal mixture-of-experts checkpoint for software agents, vision-grounded workflows and computer-use systems. The publisher reports about 428 billion total parameters with 23 billion activated per token, sparse attention throughout and support for text, image and video input with text output. Its official checkpoint uses the MiniMax Community License, not MIT or Apache; rights and practical deployment constraints need the exact license file. The newly published weights must not be conflated with an earlier proprietary hosted API evaluation. Open Model Weights source-verified the exact repository revision and license declaration before publication; it did not execute independent inference tests.
WHY IT MATTERS
A new first-party downloadable checkpoint makes a previously API-focused model family directly relevant to open-weight deployments and reproducible agent experimentation.
Sparse attention, a quantization-aware training path and a DSpark draft head make decoding and long-context behavior technically different from the existing MiniMax M3.
Computer-use and cybersecurity scores in the publisher card are claims about specific harnesses; the production value depends on a local reproducible workload.
MODEL-SELECTION VIEW
The official MiniMaxAI repository identifies a native multimodal MoE model with a newly available first-party Safetensors checkpoint. Its public documentation advertises approximately 428B logical parameters and 23B active per token. These describe distinct properties: most expert tensors still occupy storage even though sparse routing uses a subset for a specific token.
The repository includes an original model card, configuration, tokenizer and a DSpark directory. Inspecting a version-pinned file list is essential before assuming that a third-party quantization, API endpoint or earlier M3 checkpoint reproduces the same architecture. The released weights and the hosted API can evolve on different schedules.
MiniMax reports replacing the early full-attention layers with MiniMax Sparse Attention across the network. It also describes quantization-aware training for routed experts and attention paths, plus a DSpark speculative head. These changes may affect cache footprint, effective throughput and handling of unusually long sequences, but no single feature guarantees lower end-to-end agent cost.
A serving test should compare the same file revision, tokenizer, prompt structure, reasoning effort, speculative decode settings, batch size and precision. A raw tokens-per-second number without prompt length, context cache and concurrency can be misleading. Keep model-generated reasoning tokens distinct from user-visible answer tokens when measuring cost.
According to the first-party card, the model consumes text, images and video and produces text. An advertised maximum of one million tokens is a supported configuration target, not evidence that every deployment can afford the memory, achieve similar accuracy across the entire window, or decode mixed-media prompts at the same speed.
Evaluate media preprocessing, image cropping and ordering, video frame budgets, and the model's custom chat template before writing compatibility claims. A high-resolution screenshot or lengthy software repository can be truncated by the runtime long before the theoretical context capacity is reached. Use explicit task fixtures and inspect the actual prompt assembled by the client.
MiniMax publishes agent, coding and computer-use evaluations including DeepSWE, OSWorld and CyberGym. These reported values use distinct harness versions, tool configurations and budgets; they belong in an attributed publisher section, not in our independent measurement database. Review each test's success definition before drawing product conclusions.
Artificial Analysis describes its evaluated MiniMax M3.1 Flash endpoint as a proprietary API model at the time of its assessment. That is not a reason to ignore newer weight files, but the leaderboard result cannot establish that the newly downloaded checkpoint with an arbitrary local serving stack achieves the same score. Notable is the defensible initial curation tier.
The official repository declares minimax-community and distributes a LICENSE file. This is not an Apache-2.0, MIT or unrestricted public-domain declaration. Check the exact current text for attribution, commercial notification, authorization thresholds and redistribution conditions before enabling any commercial customer-facing product.
Create a manifest of the repository SHA, license text, configuration, model shards, quantization representation and serving container digest. If a license requirement is uncertain, say so rather than importing rules from the older MiniMax M3 license. Source verification must finish before canonical publication; an approved editorial profile alone never proves licensing.
The model is an attractive investigation for teams that run complex coding agents with image context, browser screenshots or lengthy tool traces. Compare it with existing canonical MiniMax M3, Ling 3.1 Flash and DeepSeek open-weight systems using the same policy sandbox, task set and abort thresholds. Use measured task success per dollar and operator workload rather than a marketing score.
A prudent rollout starts with offline reproducibility and human-reviewed agent traces, then controlled canaries. Capture prompt-injection failures, unsafe tool calls, p95 latency, memory peaks and exact checkpoint provenance. Do not claim Open Model Weights has completed inference tests unless an independently documented evaluation and saved measurement artifact exist.
IMPLEMENTATION & EVALUATION
Four-step source-pinned, reproducible evaluation plan tailored to MiniMax M3.1 Flash; not independently executed by Open Model Weights.
Record exact first-party repository SHA, LICENSE contents, shard manifest, processor and tokenizer; review custom commercial clauses.
Reproducible weights manifest and signed license assessment.
A custom-license condition can be silently missed.
Use a pinned SGLang/vLLM build and explicit text/image/video fixtures including tool parsing.
Working test prompts and exact processor logs.
Nominal one-million-token settings may exceed practical serving limits.
Evaluate code fixes, screen navigation, prompt injection resistance and independent review of tool traces.
Task-level success, failure modes and matched baseline budgets.
Publisher-selected benchmark settings may not generalize.
Profile chosen precision with realistic concurrency, latency percentiles, output budgets and actual GPU residency.
Peak VRAM, p95 response, task cost and safe rollback thresholds.
Sparse active parameters are not a memory budget.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
The official MiniMaxAI model repository lists downloadable Safetensors shards. Our source-verification gate checks the exact public file list and revision before it can enter the canonical registry.
No. The official repository declares the MiniMax Community License. Review the exact checkpoint LICENSE for commercial and attribution obligations rather than assuming MIT or Apache rights.
The publisher describes roughly 428 billion stored model parameters, while sparse experts activate around 23 billion per token. Active parameter count is not a GPU memory requirement.
MiniMax describes text, images and video as inputs, with text output. Working media pipelines still depend on supported processors and serving runtime configuration.
Not yet established here. Its cited 43-point result concerns a hosted model classified as proprietary at assessment time; it cannot be automatically transferred to a new downloadable checkpoint.
Pin revision and runtime, run representative coding and vision tool tasks with controlled budgets, then measure tool correctness, task success, peak memory, throughput and failure modes.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records MiniMax M3.1 Flash as a 220.491B-parameter text, images, video to text; agentic coding and computer use open-weight model in the MiniMax M3.1 family, with MiniMax listed as the developer. The checked source repository provides 51 recognized weight artifacts, format evidence for Safetensors and a verified context window of 1,048,576 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares MiniMax Community License. Commercial use subject to original M3.1-Flash notice, attribution and annual $20 million revenue authorization threshold. 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 subject to original M3.1-Flash notice, attribution and annual $20 million revenue authorization threshold
This section summarizes the checked repository evidence. It is not legal advice. Inspect license evidence ↗
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
dspark/model-dspark-00001.safetensorsdspark/model.safetensors.index.jsonmodel-00001-of-00048.safetensorsmodel-00002-of-00048.safetensorsmodel-00003-of-00048.safetensorsmodel-00004-of-00048.safetensorsmodel-00005-of-00048.safetensorsmodel-00006-of-00048.safetensorsmodel-00007-of-00048.safetensorsmodel-00008-of-00048.safetensorsmodel-00009-of-00048.safetensorsmodel-00010-of-00048.safetensorsmodel-00011-of-00048.safetensorsmodel-00012-of-00048.safetensorsmodel-00013-of-00048.safetensorsmodel-00014-of-00048.safetensorsmodel-00015-of-00048.safetensorsmodel-00016-of-00048.safetensorsmodel-00017-of-00048.safetensorsmodel-00018-of-00048.safetensorsmodel-00019-of-00048.safetensorsmodel-00020-of-00048.safetensorsmodel-00021-of-00048.safetensorsmodel-00022-of-00048.safetensorsmodel-00023-of-00048.safetensorsmodel-00024-of-00048.safetensorsmodel-00025-of-00048.safetensorsmodel-00026-of-00048.safetensorsmodel-00027-of-00048.safetensorsmodel-00028-of-00048.safetensorsmodel-00029-of-00048.safetensorsmodel-00030-of-00048.safetensorsmodel-00031-of-00048.safetensorsmodel-00032-of-00048.safetensorsmodel-00033-of-00048.safetensorsmodel-00034-of-00048.safetensorsmodel-00035-of-00048.safetensorsmodel-00036-of-00048.safetensorsmodel-00037-of-00048.safetensorsmodel-00038-of-00048.safetensorsmodel-00039-of-00048.safetensorsmodel-00040-of-00048.safetensorsmodel-00041-of-00048.safetensorsmodel-00042-of-00048.safetensorsmodel-00043-of-00048.safetensorsmodel-00044-of-00048.safetensorsmodel-00045-of-00048.safetensorsmodel-00046-of-00048.safetensorsmodel-00047-of-00048.safetensorsmodel-00048-of-00048.safetensorsmodel.safetensors.index.json02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
License / model-card evidence ↗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
Multimodal reasoning model. A candidate for a reasoning or coordination role. No end-to-end agent performance is inferred from its model category.
The curated notable 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: MiniMax Community License (allowed-with-conditions commercial-use classification; see original terms).
Context: 1048576 configured tokens.
BF16 weights: 440.98 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 →