Primary sources last checked on 19 Nov 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
NotableOfficial publisher original checkpoint is publicly accessible at an immutable 40-character revision, with concrete visible weights and a source-specified license. Independent OMW runtime or benchmark performance has not been established; Notable only.Curation method →
Open Model Weights records Tencent Hy4 Flash as a 220B total / ~12B active (publisher)-parameter text and image input, text output open-weight model in the Tencent Hy4 family, with Tencent Hy Team listed as the developer. The checked source repository provides 4 recognized weight artifacts, format evidence for Safetensors and a verified context window of 1,048,576 tokens, publisher-declared; not independently tested. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares Tencent HY License v1.0. Commercial use subject to Tencent HY License terms, restrictions and mandatory review. 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: publisher-described-not-reproduced; data disclosure status: Not verified from official source evidence gathered for this intake.
Tencent's public Hy4 Flash instruct checkpoint is separate from Hy4 Preview and its base/FP8 siblings. The official 220B-parameter multimodal MoE checkpoint uses about 12B active parameters per token and a publisher-declared one-million-token maximum context; it is released under the custom Tencent HY License, not Apache 2.0. The model profile links an exact original-repository revision and visible files; hardware and benchmark performance remain untested by Open Model Weights.
Official publisher sourcetencent/Hy4-flashSource ↗
Original weight evidence4 explicitly recorded source-visible Safetensors path(s); limited subset of a larger numbered shard familySource ↗
License provenancetencent-hy-license — original LICENSE fileSource ↗
Parameter and context claim220B total / ~12B active (publisher); 1,048,576 tokens, publisher-declared; not independently testedSource ↗
Benchmark statusPublisher or external leaderboard claims only; no independent OMW inference or benchmarkSource ↗
WHY IT MATTERS
The developer describes 42 language layers, sparse/sliding-window hybrid attention, LatentMoE and Engram representations, plus a vision encoder for image input.
The pinned source repository visibly exposes numbered safetensors shards named model-00001-of-00148.safetensors and onward; the source review records inspected paths, not a falsely claimed file-by-file audit of all 148 shards.
The custom Tencent HY License imposes publisher-specific conditions; record the exact license file and do not label the model generically permissive or commercially unrestricted.
MODEL-SELECTION VIEW
Strong fit when…
Source-pinned large multimodal MoE comparisons.
Long-context text and image agent research in controlled infrastructure.
Comparing the final Flash instruct weights with the separately catalogued Hy4 Preview.
Important considerations
The 1,048,576-token context is a publisher-supported capability, not a measured feasible workload on arbitrary hardware.
Weights alone do not include practical GPU memory, KV-cache and tensor-parallel costs.
The Tencent HY License requires individual review before commercial deployment or redistribution.
Model purpose and architecture
Tencent's public Hy4 Flash instruct checkpoint is separate from Hy4 Preview and its base/FP8 siblings. The official 220B-parameter multimodal MoE checkpoint uses about 12B active parameters per token and a publisher-declared one-million-token maximum context; it is released under the custom Tencent HY License, not Apache 2.0.
The developer describes 42 language layers, sparse/sliding-window hybrid attention, LatentMoE and Engram representations, plus a vision encoder for image input.
Pinned public weight artifacts
The pinned source repository visibly exposes numbered safetensors shards named model-00001-of-00148.safetensors and onward; the source review records inspected paths, not a falsely claimed file-by-file audit of all 148 shards.
The checked immutable source revision is b2498eba6a1a3f9bbcab4190bc1313a3c7010c2f. Only the explicitly listed first-party shard paths were captured in this manual review; the full numbered repository shard family was not independently enumerated. A file name is not a measured runtime result.
License and provenance
The custom Tencent HY License imposes publisher-specific conditions; record the exact license file and do not label the model generically permissive or commercially unrestricted.
License evidence: a standard original LICENSE file under tencent-hy-license. Review the complete Tencent HY-specific commercial, redistribution and acceptable-use terms.
Performance evidence and limitations
These are publisher or independently hosted third-party benchmark claims, not tests Open Model Weights has executed. A fair comparison requires an exact checkpoint, matched prompts, harness, hardware and repeatable samples.
The 1,048,576-token context is a publisher-supported capability, not a measured feasible workload on arbitrary hardware.
Deployment and practical evaluation
Appropriate evaluations include: Source-pinned large multimodal MoE comparisons. Long-context text and image agent research in controlled infrastructure. Comparing the final Flash instruct weights with the separately catalogued Hy4 Preview. Source code and quantizations from separate repositories must be checked independently.
Weights alone do not include practical GPU memory, KV-cache and tensor-parallel costs. The Tencent HY License requires individual review before commercial deployment or redistribution.
IMPLEMENTATION & EVALUATION
How to evaluate Tencent Hy4 Flash before deployment
Primary-source pinned checkpoint evaluation steps; no independent OMW model inference performed.
STEP 01
Pin official checkpoint and artifact paths
Publisher repositories can change after publication.
REPRODUCIBLE ACCEPTANCE TEST
Record the source 40-character revision and verified visible filenames, along with processor/config revision.
WHAT CAN GO WRONG
Mutable tags, missing shards and silent upstream changes.
STEP 02
Review license and code dependencies
Open model weights do not imply unrestricted use of code or all data.
REPRODUCIBLE ACCEPTANCE TEST
Inspect the exact publisher license source, custom modeling code, dependencies and intended use policy.
WHAT CAN GO WRONG
Misclassified commercial license or remote-code execution.
STEP 03
Measure hardware, context and workload
Advertised parameters, context and token rates are not observed system requirements.
REPRODUCIBLE ACCEPTANCE TEST
Publish repeatable measured memory, batch, hardware, latency and context settings.
WHAT CAN GO WRONG
Out-of-memory, unsupported attention operators and overclaimed throughput.
STEP 04
Separate external claims from OMW results
Publisher comparisons have different harnesses.
REPRODUCIBLE ACCEPTANCE TEST
Use held-out tasks and matched baselines before independent ranking.
WHAT CAN GO WRONG
False benchmarking attribution or ungrounded Leading status.
Are the original Tencent Hy4 Flash model weights publicly available?
Yes. The official source repository tencent/Hy4-flash exposes actual Safetensors checkpoint artifacts at the pinned revision b2498eba6a1a3f9bbcab4190bc1313a3c7010c2f. This review lists source-visible example weight paths, not the entire multi-shard manifest.
Is Tencent Hy4 Flash independently benchmarked by Open Model Weights?
No. Published scores and throughput descriptions are attributed to their source; OMW has not run model inference or a controlled benchmark.
What license applies to this exact checkpoint?
tencent-hy-license. Evidence is the repository's original license file; follow exact terms and separately inspect dependencies.
Is the model suitable for local deployment?
Original weights are downloadable, but actual fit, operators and performance need hardware-specific testing. Weights alone do not include practical GPU memory, KV-cache and tensor-parallel costs.
Does Notable mean the model leads an independent benchmark?
No. Notable records a distinct useful, source-verifiable original checkpoint. Leading requires separate matched independent evidence.
03 · PRACTICAL TECHNICAL PROFILE
Deployment, licensing and reproducible evidence.
Inspect weight-memory estimates, model lineage and runtime evidence separately from promotional or untested performance claims.
DEPLOYMENT PROFILE
Weight memory, without pretending it is a deployment guarantee.
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
WEIGHT-ONLY MEMORY
BF16 / FP16~440.0 GB
FP8 / INT8~220.0 GB
INT4~110.0 GB
FIELD EVIDENCE MAP
WeightsVerified
LicenseVerified
Commercial useAllowed with conditions
ContextNot declared
FormatsVerified
LineageNot declared
Training recipePublisher described not reproduced
Training dataNot verified from official source evidence gathered for this intake
RuntimeSource-derived · not runtime-tested
LICENSE & USE
Tencent HY License v1.0
Commercial use subject to Tencent HY License terms, restrictions and mandatory review
Declared licenseTencent HY License v1.0
ClassificationAllowed with conditions
VerificationLicense file checked
This section summarizes the checked repository evidence. It is not legal advice. Inspect license evidence ↗
LINEAGE & PROVENANCE
What the checked sources disclose — and what they do not.
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
LINEAGE
BASE / PARENTNot declared in model-card metadataUnknown
not inferred
CURRENT RECORDTencent Hy4 FlashTencent Hy4
TRAINING DISCLOSURE
Training recipePublisher described not reproduced
Training dataNot verified from official source evidence gathered for this intake
FamilyTencent Hy4
RUNTIME SIGNALS
Compatibility claims stay attached to their evidence level.
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
MentionedTransformers
MentionedSGLang
MentionedvLLM
OMW runtime testNot performed
EVIDENCE TRAIL
Sources and observed history.
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Three evidence-linked visual summaries: parameter scale, illustrative precision memory and verifiable provenance. No performance ranking is implied.
01 / PARAMETER SCALE
Total versus active parameters
Total parameters220B
Active parameters · publisher12B
Repository-derived or publisher-described parameter count
Total and active parameters describe different published quantities, not inference speed or runtime memory.
02 / MEMORY ESTIMATE
Weight-only precision footprint
BF16 / FP16440 GB
FP8 / INT8220 GB
INT4110 GB
Hypothetical weight-only decimal GB, not measured peak VRAM or guaranteed support for each format. KV cache, activations, runtime and sharding are excluded.
Repository SHA b2498eba6a1a…; a source check is not an inference benchmark or legal clearance. See the per-model evidence JSON for the exact source record.
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.
Declared, mentioned and externally evidenced runtimes
Declared librarytransformers
Declared/taggedNone
Mentioned in model cardTransformers, SGLang, vLLM
Artifact signalsNone
External primary runtime evidenceNone
Tested by Open Model WeightsNo — source-derived support only
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.
training recipesource-described-not-independently-reproduced
data disclosureNot verified from official source evidence gathered for this intake
runtime supportsource-derived-not-runtime-tested
MODELS FOR AI AGENTS · ROLE EVIDENCE
Where this model fits in an agent system
Specialized model. A supporting component. This checkpoint does not, by category alone, serve as the planning and tool-execution controller.
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: Tencent HY License v1.0 (allowed-with-conditions commercial-use classification; see original terms).
Context: 1048576 configured tokens.
BF16 weights: 440 estimated decimal GB, not runtime VRAM.