Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › Qwen-AgentWorld family

Qwen-AgentWorld-35B-A3B

Hugging Face ↗ Follow changes

Qwen Apache 2.0 out 4 months ago

34.7B parameters, 3B active (8 of 256 experts per token), text in, text out, 256K tokens of context.

Verified 2026-10-25 Rank #539 in LLMs · score 75.3 method v1 Sources 3 Downloads 30d 8.4k · #518 of 619 tracked

Changed this week: 2 new quantized builds on the Hub (75 in all).2026-10-24

Can I run it?21GB at INT4Estimatefits a 24 GB cardmethod
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?8.4kdownloads, 30 days#518 of 619 we track · below the median+1 likes, downloads +7% this week740 likesHugging Face2026-10-24

Worth reading

Picked by rule from 4 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 199 HN pts

Qwen-AgentWorld: Language World Models for General Agents

Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B serve as a decoupled environment simulator for scalable agentic RL and a unified agent foundation model warm-up.

arxiv.org 2026-06-24 199 HN pts · 55 comments 417 words about the Qwen AgentWorld family family
Latest · 4 months ago

Qwen-AgentWorld Models

Qwen released Qwen-AgentWorld, a native language world model trained through CPT → SFT → RL on more than 10M real environment interaction trajectories.

  • Controllable simulated RL shapes agent behavior in ways that real environments cannot and significantly outperforms RL trained solely in real-world environments.
qwen.ai 2026-06-24 3 HN pts · 1 comments 5k words about the Qwen AgentWorld family family

4 best of 4 tracked

The 3 picks above, plus these 1. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. CodeQwen-AgentWorld: Language World Models for General Agentsgithub.com · 2026-06-24 · 5 HN pts

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.

No API provider serves it, so its downloads are people running it themselves: 75 quantized builds and 6 Spaces.

  • As a service0no API provider listed
  • On own hardware75quantized builds
  • Built on2311 fine-tunes · 2 adapters · 10 merges
  • In apps6Hugging Face Spaces

Top Spaces running it: chabab/qwen-agentworld-35b-zerogpu ♥ 3 · geraldamasi/agentworld-demo ♥ 2

Will it fit?

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

Fits a 64 GB Mac (unified memory) at INT4

About 21 GB at INT4, 42 GB at FP8, 83 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 21 GB of weights (estimate; KV cache and activations extra)INT421 GBFP8: about 42 GB of weights (estimate; KV cache and activations extra)FP842 GBBF16: about 83 GB of weights (estimate; KV cache and activations extra)BF1683 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 21 GB of weights (estimate; KV cache and activations extra)INT421 GBFP8: about 42 GB of weights (estimate; KV cache and activations extra)FP842 GBBF16: about 83 GB of weights (estimate; KV cache and activations extra)BF1683 GB

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 (2026-06-22), amber triangles are 4 dated articles (each links to it), grey ticks are repo commits.

JulAugSepOct2026-06-25: 1 repo commit2026-06-24: 1 repo commit2026-06-23: 6 repo commits2026-06-22: 1 repo commitQwen-AgentWorld: Language World Models for General Agents · 2026-06-24Qwen-AgentWorld Models · 2026-06-24Qwen-AgentWorld: Language World Models for General Agents · 2026-06-24Qwen-AgentWorld-35B-A3B: a local 'world model' you can run at home Open Models · 2026-06-27Weights 2026-06-22 · 2026-06-22Weights 2026-06-22now
JulAugSepOct2026-06-25: 1 repo commit2026-06-24: 1 repo commit2026-06-23: 6 repo commits2026-06-22: 1 repo commitQwen-AgentWorld: Language World Models for General Agents · 2026-06-24Qwen-AgentWorld Models · 2026-06-24Qwen-AgentWorld: Language World Models for General Agents · 2026-06-24Qwen-AgentWorld-35B-A3B: a local 'world model' you can run at home Open Models · 2026-06-27Weights 2026-06-22 · 2026-06-22Weights 2026-06-22now

What people built on it

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

98 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.

75 quantized (Hugging Face model tree)11 fine-tunes (Hugging Face model tree)10 merges (Hugging Face model tree)2 adapters (Hugging Face model tree)75 quantized11 fine-tunes10 merges2 adapters
75 quantized (Hugging Face model tree)11 fine-tunes (Hugging Face model tree)10 merges (Hugging Face model tree)2 adapters (Hugging Face model tree)75 quantized11 fine-tunes10 merges2 adapters
GGUF unsloth/Qwen-AgentWorld-35B-A3B-GGUF for llama.cpp, Ollama, LM Studio 1.9M downloads MLX mlx-community/Qwen-AgentWorld-35B-A3B-oQ4 for Apple silicon 4.4k downloads FP8 lovedheart/Qwen-AgentWorld-35B-A3B-FP8 for vLLM, SGLang 16.4k downloads

All variants →

Reference

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

Benchmarks

8 rows

Benchmarks 8 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (8)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.
BenchmarkValueSetting / notesReported byDate
author-reported AgentWorldBench64.79MCPmodel author2026-10-07
author-reported AgentWorldBench36.69Searchmodel author2026-10-07
author-reported AgentWorldBench53.96Term.model author2026-10-07
author-reported AgentWorldBench65.63SWEmodel author2026-10-07
author-reported AgentWorldBench58.17Androidmodel author2026-10-07
author-reported AgentWorldBench49.55Webmodel author2026-10-07
author-reported AgentWorldBench65.92OSmodel author2026-10-07
author-reported AgentWorldBench56.39Overallmodel author2026-10-07

What the card says

2 cited sentences

Summary generated, cited

Qwen-AgentWorld-35B-A3B is a mixture-of-experts language world model designed for agentic environment simulation.[1][2] It is released under the Apache 2.0 license and supports a context length of 262,144 tokens.[3][4]

  1. is mixture of experts
  2. “native language world model trained for agentic environme…”
  3. license
  4. “default context length of 262,144 tokens”

Generated 2026-10-07 15:12 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

6 use cases

What it is good at 6 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

24 fields

Listed in LLMs · MoE

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesQwen3_5MoeForConditionalGeneration
Model typeqwen3_5_moe
Dtypebfloat16
Params by dtypeBF16: 34,660,610,688
Size bandL
MoEyes
Experts256
Experts per token8
Context length262,144 (config.json max_position_embeddings)
PapersarXiv:2606.24597
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Qwen3_5MoeForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen3_5MoeForConditionalGeneration · 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 Qwen-AgentWorld-35B-A3B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)83 GB · Multi-GPU node (<640 GB)
FP8 (1 byte/param)42 GB · Single high-end GPU (<80 GB)
INT4 (0.5 bytes/param)21 GB · Single GPU (<24 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

Fine-tuned from Qwen/Qwen3.5-35B-A3B-Base (same organization) Family: Qwen/Qwen-AgentWorld 98 derivatives on the Hub

Base model: Qwen/Qwen3.5-35B-A3B-Base

quantized: 75merge: 10finetune: 11adapter: 2

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
unsloth/Qwen-AgentWorld-35B-A3B-GGUF quantized GGUF — — apache-2.0 unsloth 251 1.9M
mudler/Qwen-AgentWorld-35B-A3B-APEX-GGUF quantized GGUF — — apache-2.0 mudler 13 37.4k
SC117/Qwen-AgentWorld-35B-A3B-MTP-Uncensored-APEX-GGUF quantized GGUF — — apache-2.0 SC117 14 37.2k
DevQuasar/Qwen.Qwen-AgentWorld-35B-A3B-GGUF quantized GGUF — — — DevQuasar 1 20.3k
SC117/Qwen-AgentWorld-35B-A3B-MTP-APEX-GGUF quantized GGUF — — apache-2.0 SC117 8 19.1k
lovedheart/Qwen-AgentWorld-35B-A3B-FP8 quantized -FP8 34.7B L apache-2.0 lovedheart 3 16.4k
groxaxo/Qwen-AgentWorld-35B-A3B-GGUF quantized GGUF — — — groxaxo 4 15.4k
cyankiwi/Qwen-AgentWorld-35B-A3B-AWQ-INT4 quantized AWQ 34.7B L apache-2.0 cyankiwi 4 15.1k
pragmaticcs/Qwen-35B-A3B-SignOfFour-Coder-GGUF merge GGUF — — apache-2.0 pragmaticcs 8 12.7k
FreedomAISVR/Qwen-AgentWorld-35B-A3B-NVFP4-GGUF quantized NVFP4 — — apache-2.0 FreedomAISVR 3 10.6k
DhruvalLabs/Qwen-AgentWorld-35B-A3B-GGUF quantized GGUF — — apache-2.0 DhruvalLabs 6 8.2k
Jiunsong/SuperQwen-AgentWorld-35B-A3B-abliterated finetune — 34.7B L apache-2.0 Jiunsong 19 7.3k
gaoqianshen/Qwen-AgentWorld-35B-A3B-Q4_K_M-GGUF quantized GGUF — — apache-2.0 gaoqianshen 4 6.5k
lovedheart/Qwen-AgentWorld-35B-A3B-NVFP4 quantized NVFP4 21.7B L apache-2.0 lovedheart 10 5.6k
FreedomAISVR/Qwen-AgentWorld-35B-A3B-MXFP4-MOE-GGUF quantized MXFP4 — — apache-2.0 FreedomAISVR 2 5.8k
mlx-community/Qwen-AgentWorld-35B-A3B-oQ4 quantized — 34.7B L apache-2.0 mlx-community 12 4.4k
ivgranite/Qwen-AgentWorld-35B-A3B-GGUF quantized GGUF — — apache-2.0 ivgranite 1 4.4k
hotdogs/Qwen35B-Agent-R2 quantized GGUF 34.7B L agpl-3.0 hotdogs 6 4k
vadikulous-ai/Qwen-AgentWorld-35B-A3B-UD-Q4_K_XL-MTP-GGUF quantized GGUF — — apache-2.0 vadikulous-ai 5 3.6k
sakamakismile/Qwen-AgentWorld-35B-A3B-NVFP4 quantized NVFP4 20.4B L apache-2.0 sakamakismile 8 3.3k
OliviaRossi/UltiMerge merge — 34.7B L apache-2.0 OliviaRossi 10 3.1k
pragmaticcs/Qwen-35B-A3B-SignOfFour-Coder merge — 35.1B XL apache-2.0 pragmaticcs 6 3.3k
huihui-ai/Huihui-Qwen-AgentWorld-35B-A3B-abliterated finetune — 34.7B L apache-2.0 huihui-ai 19 2.2k
jedisct1/Qwen-AgentWorld-35B-A3B-oQ4-MLX quantized MLX 34.7B L apache-2.0 jedisct1 2 2.6k
cyankiwi/Qwen-AgentWorld-35B-A3B-AWQ-FP8 quantized AWQ 34.7B L apache-2.0 cyankiwi 1 2.1k
unsloth/Qwen-AgentWorld-35B-A3B finetune — 34.7B L apache-2.0 unsloth 5 1.9k
mlx-community/Qwen-AgentWorld-35B-A3B-bf16 finetune -bf16 34.7B L apache-2.0 mlx-community 3 2k
havok2/Qwen-AgentWorld-35B-A3B-VL36 merge — 35.1B XL apache-2.0 havok2 4 1.9k
groxaxo/Qwen-AgentWorld-35B-A3B-GPTQ-Pro-Int4 quantized GPTQ 34.7B L apache-2.0 groxaxo 3 1.8k
auttasak88/AgentWorld-35B-A3B-Heretic finetune GGUF 35.1B XL apache-2.0 auttasak88 1 1.8k
OsaurusAI/Qwen-AgentWorld-35B-A3B-MXFP8 finetune — 34.7B L apache-2.0 OsaurusAI 5 1.3k
jedisct1/Qwen-AgentWorld-35B-A3B-oQ8-MLX quantized MLX 34.7B L apache-2.0 jedisct1 1 1.4k
AtomicChat/qwen-agentworld-35b-GGUF quantized GGUF — — apache-2.0 AtomicChat 1 1.4k
DreamFoundries/Qwen-AgentWorld-35B-A3B-6bit quantized — 34.7B L apache-2.0 DreamFoundries 0 1.4k
pragmaticcs/Qwen3.8-KAT-AgentWorld-35B-A3B-Coder-GGUF merge GGUF — — apache-2.0 pragmaticcs 10 884
Ichigec/SuperQwen-AgentWorld-35B-A3B-abliterated-APEX-I-Quality quantized GGUF — — apache-2.0 Ichigec 0 1.3k
jedisct1/Qwen-AgentWorld-35B-A3B-REAP78-oQ4-MLX quantized MLX 9.48B L apache-2.0 jedisct1 3 1.2k
Aminlight/Qwen-AgentWorld-35B-A3B-4bit-MLX quantized -4bit 34.7B L apache-2.0 Aminlight 2 1.2k
LeaderboardModel1/Qwen-AgentWorld-35B-A3B-AutoRound-W4A16-RTN quantized -W4A16 1.49B S — LeaderboardModel1 2 1.2k
groxaxo/Qwento-Agentic finetune — 34.7B L apache-2.0 groxaxo 5 907

Repository changelog

9 commits
changed in the last 7 daysOther: 2026-10-24all changes

Repository changelog 9 commits

DateCommitSHA
2026-06-25 Update README.md (#5) 60d2b04
2026-06-24 Update README.md fec3dfb
2026-06-23 Update README.md 6de136d
2026-06-23 Update README.md 732a3e9
2026-06-23 Delete .mv 787144d
2026-06-23 Delete .msc 0374630
2026-06-23 Update README.md 303557d
2026-06-23 Upload folder using huggingface_hub a6a5fe5
2026-06-22 initial commit d4107a8

Articles and coverage

4 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

Qwen-AgentWorld Models
qwen.ai2026-06-24officialauthorityabout the Qwen AgentWorld familyHN 3 pts · 1 comments4,975 words
Qwen released Qwen-AgentWorld, a native language world model trained through CPT → SFT → RL on more than 10M real environment interaction trajectories. The model functions as a decoupled environment simulator for scalable agentic RL and as a unified agent foundation model where LWM warm-up enables effective transfer to multi-turn agentic tasks across seven benchmarks, including three entirely out of domain. This single model simulates seven domains including MCP, Search, Terminal, SWE, Web, OS, and Android.written, reviewed, cited
  • Controllable simulated RL shapes agent behavior in ways that real environments cannot and significantly outperforms RL trained solely in real-world environments.

Analysis3

Qwen-AgentWorld: Language World Models for General Agents
arxiv.org2026-06-24authorityHN front pageabout the Qwen AgentWorld familyHN 199 pts · 55 comments417 words
Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B serve as a decoupled environment simulator for scalable agentic RL and a unified agent foundation model warm-up. Leveraging more than 10M environment interaction trajectories of 7 domains, these models support thousands of real-world environment simulations and improve downstream performance across 7 agentic benchmarks. The architecture implies MoE structures via A3B and A17B designations, enabling engineers to assess compute requirements for general agent tasks.written, reviewed, cited
Qwen-AgentWorld: Language World Models for General Agents
github.com2026-06-24authorityabout the Qwen AgentWorld familyHN 5 pts · 1 comments1,641 words
Qwen-AgentWorld-35B-A3B is a language world model released on 2026-06-24 that simulates agentic environments across seven unified domains using a mixture-of-experts architecture with 35B total / 3B active parameters and 256K context. A larger 397B variant achieves the highest overall score of 58.71, outperforming frontier proprietary models like GPT-5.4 on the benchmark. The model is accompanied by the AgentWorldBench evaluation benchmark which scores predictions on five dimensions including Format, Factuality, Consistency, Realism, and Quality.written, reviewed, cited
  • The vLLM inference command requires the --language-model-only flag because the checkpoint contains only language weights despite visual component definitions.
  • The evaluation pipeline uses a specific judge model identified as gpt-5.2-2025-12-11 to score the world model's generated observations.
  • Each sample in the AgentWorldBench dataset carries its own system prompt in the system_str field rather than using a single global template.
Qwen-AgentWorld-35B-A3B: a local 'world model' you can run at home Open Models
vettedconsumer.com2026-06-27HN 2 pts · 0 comments1,617 words
Qwen released Qwen-AgentWorld-35B-A3B on June 22 as an Apache-2.0 licensed local language world model that predicts environment states rather than picking actions. The 35 billion parameter model requires approximately 21 GB of VRAM at Q4_K_M quantization, necessitating a 24 GB card or unified memory machine to hold all weights before generating. It covers seven agent domains including OS and SWE, scoring 56.39 on the AgentWorldBench benchmark.written, reviewed, cited
  • The model uses Gated DeltaNet blocks and gated attention to keep KV cache growth in check during long-context runs.
  • All 35 billion parameters must be loaded into memory first, requiring a 24 GB card despite only 3 billion being active.
  • The technical report was submitted on June 23, 2026, which is after the model's June 22 release date.

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Similar models

3 models

Similar models

  • GLM-4.7-Flash · 31.2B
    same size band (L), also MoE, same task (text-generation), more providers, permissive license
  • Qwen3-Coder-30B-A3B-Instruct · 30.5B
    same size band (L), also MoE, same task (text-generation), more providers, permissive license
  • Qwen3-30B-A3B · 30.5B
    same size band (L), also MoE, same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:40 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:12 UTC · Last verified 2026-10-25

Follow Qwen-AgentWorld-35B-A3B

Watch this model

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