Hy3

299B parameters (8 of 192 experts per token), text in, text out, 256K tokens of context. 7 providers serve it from $0.083 per million input tokens.

Verified 2026-10-25 Rank #149 in LLMs · score 93.1 method v1 Sources 4 Downloads 30d 941.7k · #118 of 619 tracked

Changed this week: Cheapest input price fell 36.5%: $0.130 → $0.083 per 1M (OpenRouter).2026-10-24 · 1 new quantized build on the Hub (66 in all).2026-10-24

Can I run it?179GB at INT4Estimateneeds a multi-GPU nodemethodCommunity: Tencent Hy3: 74.4% SWE-bench With Only 21B Active Parameters (2026) →

What does it cost?$0.083/1M incheapest of 6Tencentfp82026-10-24

Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24

Do people use it?941.7kdownloads, 30 days#118 of 619 we track · top 19%+5 likes, downloads +31% this week987 likesHugging Face2026-10-24

Worth reading

Picked by rule from 15 tracked pieces: the lab's own post, the most shared, a field report, the latest.

The lab's own post

Hy3

Tencent released Hy3 on July 6, 2026, as an updated version of its Hunyan model available under the Apache 2.0 license.

hy.tencent.com 2026-07-07 561 HN pts · 120 comments 197 words sells inference

Most discussed · 150 HN pts

The mysterious Hy3 LLM is topping OpenRouter Model Rankings by a large margin

Hy3 preview from Tencent tops OpenRouter rankings despite DeepSeek V4 Flash offering significantly lower effective pricing of $0.018/1M input tokens via a 2% cache read cost.

minimaxir.com 2026-05-26 150 HN pts · 112 comments 1.6k words

Latest · 2 months ago

OpenCode go provides Hy3 with 8X usage

OpenCode Go is a $10 per month subscription providing reliable access to curated coding models for agentic workflows.

Muse Spark 1.3 and 1.2 Contributor models require permission to use prompts for training future Meta models.
Grok 4.7 disables the stateful Responses API, Files and Collections, and the Batch API due to ZDR.

opencode.ai 2026-08-21 2 HN pts 1k words

15 best of 15 tracked

The 4 picks above, plus these 11. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

DiscussionHy3 – 295B MoE model (21B active, 192 experts)twitter.com · 2026-07-07 · 2 HN pts
ArticleWhat Is Hy3? Tencent's 295B Open-Source MoEkie.ai · 2026-07-16
ArticleTencent's Hy3 Bets On AI Agents Over Model Sizeforbes.com · 2026-07-13
ArticleTencent Hy3: 74.4% SWE-bench With Only 21B Active Parameters (2026)aimadetools.com · 2026-07-11
ArticleTencent Hunyuan Hy3 Deep Dive: 295B MoE Fast-Slow Thinking Fusion, Apache 2.0 Open Source, 90% Agent Task Success Rate | HappyRockhappyrock.cloud · 2026-07-08
ArticleTencent Hy3: 295B Params, 21B Active — Can You Run It? - DEV Communitydev.to · 2026-07-08
ArticleTencent Hunyuan Hy3: Does Smaller Actually Beat Flagship Open Weights · Groundygroundy.com · 2026-07-08
ArticleTencent Hunyuan Hy3 Developer Guide & Benchmarks | Lushbinarylushbinary.com · 2026-07-06
ArticleTencent's Hunyuan Hy3: Open-Weight Reasoning Arrivesdigitalapplied.com · 2026-07-05
ArticleTencent used Anthropic's Claude to fine-tune it's new Hy3 AI modelreuters.com · 2026-04-29
ArticleTencent Hy3 Preview: The 295B Open MoE That Topped OpenRouter — and What Builders Should Actually Know — ChatForestchatforest.com · 2026-06-14

Show all 11

Where to run it

Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.

Prices for the same weights vary 2×

6 providers with a live price, from $0.083 to $0.20 per million input tokens; the median is $0.14. One dot per provider, cheapest in teal.

#	Cheapest 5	In /1M	Out /1M	Context	Quant	Seen
1	Tencent	$0.083	$0.33	256K	fp8	Oct 24
2	DeepInfra	$0.13	$0.53	256K	—	Oct 25
3	GMICloud	$0.14	$0.58	256K	bf16	Oct 24
4	Novita	$0.14	$0.58	256K	unknown	Oct 24
5	Phala	$0.15	$0.64	256K	unknown	Oct 24

All 7 providers →

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

Widely served (7 API providers); heavily run locally (66 quantized builds).

As a service7AtlasCloud, DeepInfra, Featherless, GMICloud +3
On own hardware66quantized builds
Built on1313 fine-tunes · 0 adapters · 0 merges
In apps27Hugging Face Spaces

Top Spaces running it: DeepImagix/self-trained2 ♥ 1 · akhaliq/Hy3-new ♥ 1 · vora-sonnet/aurora-ai-os ♥ 1 · abonaie222263/hf-openai-proxy ♥ 1

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Needs an 8×80 GB node even at INT4

About 179 GB at INT4, 359 GB at FP8, 717 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Most coverage came in the first two weeks

Teal is the weights (2026-07-02), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

What people built on it

79 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

79 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

GGUF AngelSlim/Hy3-GGUF for llama.cpp, Ollama, LM Studio 1.2M downloads MLX avlp12/Hy3-Alis-MLX-Dynamic for Apple silicon 5.3k downloads NVFP4 kodelow/Hy3-NVFP4-W4A16 for vLLM, SGLang 9.7k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

1 row

Benchmarks 1 row

measured third-party leaderboard rowauthor-reported number from the model card or eval files (1)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.

Benchmark	Value	Setting / notes	Reported by	Date
author-reported blind evaluation with 270 experts	2.67/4	—	model author	2026-10-07

Use-case evidence

4 use cases

What it is good at 4 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

22 fields

Listed in LLMs · MoE

Facts

Hy3 has 299B parameters. (safetensors metadata, observed 2026-10-24)
Hy3 is a mixture-of-experts model with 192 experts, routing 8 per token. (config.json, observed 2026-10-24)
Context length is 262,144 tokens (256K). (config.json max_position_embeddings, observed 2026-10-24)
License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
Created 2026-07-02, last modified 2026-07-20. (HF API, observed 2026-10-24)

Pipeline tag	text-generation
Library	transformers
Architectures	HYV3ForCausalLM
Model type	hy_v3
Dtype	bfloat16
Params by dtype	BF16: 298,786,140,416 · F32: 15,360
Size band	XXL
MoE	yes
Experts	192
Experts per token	8
Context length	262,144 (config.json max_position_embeddings)
License	Apache 2.0 (apache-2.0)
Hub snapshot	2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · HYV3ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Hy3: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)	717 GB · Cluster (≥640 GB)
FP8 (1 byte/param)	359 GB · Multi-GPU node (<640 GB)
INT4 (0.5 bytes/param)	179 GB · Multi-GPU node (<640 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

40 variants

Lineage

Original model Family: tencent/Hy3 79 derivatives on the Hub

quantized: 66finetune: 13

Top variants (40 shown, ranked by downloads)

Variant	Relation	Quant	Params	Band	License	Author	Likes	Downloads
AngelSlim/Hy3-GGUF	quantized	GGUF	—	—	apache-2.0	AngelSlim	187	1.2M
bartowski/Hy3-GGUF	quantized	GGUF	—	—	apache-2.0	bartowski	10	95.2k
YanissAmz/Hy3-295B-A21B-GGUF	quantized	GGUF	—	—	apache-2.0	YanissAmz	21	54.6k
satgeze/Hy3-1M-GGUF	quantized	GGUF	—	—	apache-2.0	satgeze	45	33.2k
giannisan/Hy3-ds4-gguf	quantized	gguf	—	—	—	giannisan	30	19.8k
kodelow/Hy3-NVFP4-W4A16	quantized	NVFP4	172B	XXL	apache-2.0	kodelow	15	9.7k
rdtand/Hy3-295B-A21B-PrismaQuant-2.8bit-gguf-vllm	quantized	gguf	—	—	apache-2.0	rdtand	2	5.8k
avlp12/Hy3-Alis-MLX-Dynamic	quantized	MLX	295B	XXL	apache-2.0	avlp12	10	5.3k
RedHatAI/Hy3-NVFP4-FP8	quantized	NVFP4	299B	XXL	mit	RedHatAI	7	5.3k
AngelSlim/Hy3-GPTQ-Int4	quantized	GPTQ	299B	XXL	apache-2.0	AngelSlim	10	4.5k
olka-fi/Hy3-MXFP4	quantized	MXFP4	163B	XXL	apache-2.0	olka-fi	6	4.7k
mlx-community/Hy3-oQ2	quantized	—	295B	XXL	apache-2.0	mlx-community	9	4.1k
jcbtc/Hy3-Chadrock-FPX-IFP2-MTP	quantized	GGUF	—	—	apache-2.0	jcbtc	3	4.3k
cyankiwi/Hy3-AWQ-INT4	quantized	AWQ	49.6B	XL	apache-2.0	cyankiwi	4	4.2k
cyankiwi/Hy3-AWQ-NVFP4	quantized	AWQ	167B	XXL	apache-2.0	cyankiwi	2	4.2k
LibertAIDAI/Hy3-NVFP4	quantized	NVFP4	172B	XXL	other	LibertAIDAI	2	3.9k
ox-ox/Hy3-295B-Instruct-w2q3exp-AProjQ8-SExpQ8-OutQ8-MTP-mlx	quantized	mlx	—	—	apache-2.0	ox-ox	3	3k
Jiunsong/SuperHY3-abliterated-NVFP4	finetune	NVFP4	172B	XXL	apache-2.0	Jiunsong	16	2.1k
mlx-community/Hy3-oQ2e	quantized	—	295B	XXL	apache-2.0	mlx-community	4	2.7k
canada-quant/hy3-w4a16-mtp	quantized	-w4a16	46.8B	XL	apache-2.0	canada-quant	5	2.6k
LordNeel/Hy3-GGUF	quantized	GGUF	—	—	apache-2.0	LordNeel	4	2.3k
mlx-community/Hy3-OptiQ-2bit	quantized	—	295B	XXL	apache-2.0	mlx-community	6	2k
AtomicChat/Hy3-GGUF	quantized	GGUF	—	—	apache-2.0	AtomicChat	5	2k
philipjohnbasile/hy3-demolition-mlx-lite-v1	quantized	mlx	295B	XXL	apache-2.0	philipjohnbasile	1	1.8k
tiny-random/hy-v3	finetune	—	2.57M	S	—	tiny-random	1	1.8k
philipjohnbasile/hy3-demolition-mlx-reap25-v1	quantized	mlx	223B	XXL	apache-2.0	philipjohnbasile	2	1.7k
rdtand/Hy3-295B-A21B-PrismaQuant-5.3bit-2xSpark-vllm	quantized	—	194B	XXL	apache-2.0	rdtand	1	1.7k
yujiepan/hy-v3-tiny-random	finetune	—	2.57M	S	—	yujiepan	0	1.7k
JANGQ-AI/Hy3-JANG_2K-MTP	quantized	—	299B	XXL	other	JANGQ-AI	4	1.5k
INCModel2/Hy3-MXFP4-Mixed-CT-AutoRound-Preview	quantized	MXFP4	163B	XXL	other	INCModel2	1	1.6k
yujiepan/hy3-tiny-random	finetune	—	2.57M	S	—	yujiepan	0	1.6k
tiyuvta/Hy3-NVFP4	quantized	NVFP4	154B	XXL	apache-2.0	tiyuvta	1	1.5k
pipenetwork/Hy3-REAP75-MLX-4bit	quantized	MLX	80.3B	XL	apache-2.0	pipenetwork	4	1.2k
cloudyu/hy3-gguf	quantized	gguf	—	—	apache-2.0	cloudyu	1	1.1k
philipjohnbasile/hy3-demolition-mlx-reap25-v1-mtp	quantized	mlx	223B	XXL	apache-2.0	philipjohnbasile	2	971
unigilby/Hy3-oQ4e	quantized	—	295B	XXL	apache-2.0	unigilby	1	978
pipenetwork/Hy3-REAP62-MLX-4bit	quantized	MLX	116B	XXL	apache-2.0	pipenetwork	3	857
OpensourceWTF/Hy3-oQ2e-MTPLX-streaming	quantized	—	8.75B	M	apache-2.0	OpensourceWTF	3	787
pipenetwork/Hy3-2bit	quantized	—	295B	XXL	apache-2.0	pipenetwork	3	771
vroomfondel/Hy3-NVFP4-W4A4	quantized	NVFP4	148B	XXL	apache-2.0	vroomfondel	1	855

Repository changelog

15 commits

changed in the last 7 daysPrice moves: 2026-10-24Other: 2026-10-24all changes

Repository changelog 15 commits

Date	Commit	SHA
2026-07-20	Update config.json	a960ebc
2026-07-17	Upload folder using huggingface_hub	02b95e9
2026-07-16	Upload folder using huggingface_hub	89ffd53
2026-07-16	Upload folder using huggingface_hub	e4e6605
2026-07-06	Upload folder using huggingface_hub	716aa72
2026-07-06	Update README.md	6acfeb6
2026-07-06	Update README_CN.md	24fe753
2026-07-06	Update vllm recipes and sglang codebook	45b92d8
2026-07-06	Upload folder using huggingface_hub	62b972e
2026-07-06	Upload folder using huggingface_hub	5137deb
2026-07-05	Upload chat_template.jinja with huggingface_hub	3917583
2026-07-04	Add files using upload-large-folder tool	9feb605
2026-07-04	Add files using upload-large-folder tool	d4bf2b3
2026-07-04	Add files using upload-large-folder tool	c1131ba
2026-07-02	initial commit	bee39e3

All providers

7 providers · 7 endpoints

Where to run 7 providers · 7 endpoints

Hosted endpoints observed for Hy3, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.08/M.

Provider	Model ref	Context	Quant	$/M in	$/M out	tok/s	TTFT	Tools	Status	Observed
Tencent openrouter	tencent/hy3	256K	fp8	$0.08	$0.33	—	—	yes	live	2026-10-24
DeepInfra hf	tencent/Hy3	256K	—	$0.13	$0.53	52.1	674 ms	yes	live	2026-10-25
Novita openrouter	tencent/hy3	256K	unknown	$0.14	$0.58	—	—	yes	live	2026-10-24
GMICloud openrouter	tencent/hy3	256K	bf16	$0.14	$0.58	—	—	yes	live	2026-10-24
Phala openrouter	tencent/hy3	256K	unknown	$0.15	$0.64	—	—	yes	live	2026-10-24
AtlasCloud openrouter	tencent/hy3	256K	fp8	$0.20	$0.80	—	—	yes	live	2026-10-24
Featherless hf	tencent/Hy3	—	—	—	—	—	—	—	live	2026-10-25

Articles and coverage

15 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official1

Hy3

hy.tencent.com2026-07-07officialHN front pageHN 561 pts · 120 comments197 words

Tencent released Hy3 on July 6, 2026, as an updated version of its Hunyan model available under the Apache 2.0 license. Internal testing shows the model reduced hallucination rates to 5.4% and increased task resolution to 90% in WorkBuddy. Users report significant efficiency gains, with document processing saving 47.4% in token consumption compared to GLM-5.2.written, reviewed, cited

Analysis8

The mysterious Hy3 LLM is topping OpenRouter Model Rankings by a large margin

minimaxir.com2026-05-26HN front pageHN 150 pts · 112 comments1,633 words

Hy3 preview from Tencent tops OpenRouter rankings despite DeepSeek V4 Flash offering significantly lower effective pricing of $0.018/1M input tokens via a 2% cache read cost. Hy3 costs $0.066/1M tokens input via SiliconFlow with a restrictive license that could prevent providers from adopting the model. Usage data shows a 98% input, 2% output token breakdown since switching from free to paid status on May 8.written, reviewed, cited

Hy3 (295B A21B), a reasoning and agent model with great cost efficiency

github.com2026-07-06authorityHN 6 pts · 0 comments963 words

Tencent released Hy3, a 295B-parameter Mixture-of-Experts model with 21B active parameters and a 256K context length designed for high cost efficiency. The model reduced its hallucination rate from 12.5% to 5.4% through fine-grained data cleaning and is released under the Apache License 2.0. It outperforms similar-size open-source models and rivals flagship models with 2-5x parameters while showing significant gains in utility across various productivity tasks.written, reviewed, cited

Hy3 requires H20-3e or GPUs with larger memory capacity to serve its 295B parameters on 8 GPUs.
Hy3 supports GRPO reinforcement learning training via verl or Megatron-LM conversion using NVIDIA Megatron-Bridge.

OpenCode go provides Hy3 with 8X usage

opencode.ai2026-08-21HN 2 pts · 0 comments1,027 words

OpenCode Go is a $10 per month subscription providing reliable access to curated coding models for agentic workflows. Go Plus costs $40 per month with higher limits, while quotas enforce 20% of the monthly allowance per 5 hours. Some models like Grok 4.7 retain data for 30 days, whereas others retain zero days.written, reviewed, cited

Muse Spark 1.3 and 1.2 Contributor models require permission to use prompts for training future Meta models.
Grok 4.7 disables the stateful Responses API, Files and Collections, and the Batch API due to ZDR.
DeepSeek V4.1 Flash operates under a ZDR agreement valid through October 31, 2026.

What Is Hy3? Tencent's 295B Open-Source MoE

kie.ai2026-07-161,864 words

Tencent's Hy3 Bets On AI Agents Over Model Size

forbes.com2026-07-13

Tencent Hy3: 74.4% SWE-bench With Only 21B Active Parameters (2026)

aimadetools.com2026-07-111,553 words

Tencent Hunyuan Hy3 Deep Dive: 295B MoE Fast-Slow Thinking Fusion, Apache 2.0 Open Source, 90% Agent Task Success Rate | HappyRock

happyrock.cloud2026-07-081,030 words

Tencent Hy3: 295B Params, 21B Active — Can You Run It? - DEV Community

dev.to2026-07-081,396 words

Community1

Hy3 – 295B MoE model (21B active, 192 experts)

twitter.com2026-07-07HN 2 pts · 1 comments

All coverage tracked for this model (15) — append-only, nothing is ever removed

Title	Source	Published	First seen	Signals
OpenCode go provides Hy3 with 8X usage	opencode.ai	2026-08-21	2026-10-07	hn
What Is Hy3? Tencent's 295B Open-Source MoE	kie.ai	2026-07-16	2026-10-07	exa
Tencent's Hy3 Bets On AI Agents Over Model Size	forbes.com	2026-07-13	2026-10-07	exa
Tencent Hy3: 74.4% SWE-bench With Only 21B Active Parameters (2026)	aimadetools.com	2026-07-11	2026-10-07	exa
Tencent Hunyuan Hy3 Deep Dive: 295B MoE Fast-Slow Thinking Fusion, Apache 2.0 Open Source, 90% Agent Task Success Rate \| HappyRock	happyrock.cloud	2026-07-08	2026-10-07	exa
Tencent Hy3: 295B Params, 21B Active — Can You Run It? - DEV Community	dev.to	2026-07-08	2026-10-07	exa
Tencent Hunyuan Hy3: Does Smaller Actually Beat Flagship Open Weights · Groundy	groundy.com	2026-07-08	2026-10-07	exa
Hy3 – 295B MoE model (21B active, 192 experts)	twitter.com	2026-07-07	2026-10-07	hn
Hy3	hy.tencent.com official	2026-07-07	2026-10-07	hn
Hy3 (295B A21B), a reasoning and agent model with great cost efficiency	github.com	2026-07-06	2026-10-07	hn
Tencent Hunyuan Hy3 Developer Guide & Benchmarks \| Lushbinary	lushbinary.com	2026-07-06	2026-10-07	exa
Tencent's Hunyuan Hy3: Open-Weight Reasoning Arrives	digitalapplied.com	2026-07-05	2026-10-07	exa
Tencent Hy3 Preview: The 295B Open MoE That Topped OpenRouter — and What Builders Should Actually Know — ChatForest	chatforest.com	2026-06-14	2026-10-07	exa
The mysterious Hy3 LLM is topping OpenRouter Model Rankings by a large margin	minimaxir.com	2026-05-26	2026-10-07	hn
Tencent used Anthropic's Claude to fine-tune it's new Hy3 AI model	reuters.com	2026-04-29	2026-10-07	hn

Similar models

3 models

Similar models

DeepSeek-V4-Flash-0731 · 304B
same size band (XXL), also MoE, same task (text-generation), more providers, permissive license
DeepSeek-V4-Flash · 291B
same size band (XXL), also MoE, same task (text-generation), more providers, permissive license
MiMo-V2.6-Flash-RL · 311B
same size band (XXL), also MoE, same task (text-generation), fewer providers, permissive license

Sources

4 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:35 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:03 UTC · Last verified 2026-10-25

Follow Hy3

Watch this model

/feeds/models/tencent--Hy3.xml(RSS 2.0) · every change to Hy3
/feeds/families/tencent--Hy3.xml(RSS 2.0) · every change across the Hy3 family

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.