Sign in

Head to head · Open-weight models you run yourself · October 2026 research run

Qwen3-30B-A3B-Instruct-2507 vs WeVisOne-VL-4B

WeVisOne-VL-4B scores 61 (C) on agent readiness against Qwen3-30B-A3B-Instruct-2507's 56.6 (C), and leads in 1 of 4 scored categories. Qwen3-30B-A3B-Instruct-2507 leads on reach and agent interface. Both do open-weight models you run yourself.

Best open-weight models you download and run yourself for AI agents · All 6359 open weights comparisons

Qwen3-30B-A3B-Instruct-2507Alibaba (Qwen team) · Open-weight model C56.6/100#113 of 160 open-weight models WeVisOne-VL-4BTencent (WeChat Vision) · Open-weight model C61/100#84 of 160 open-weight models

Which one, for what

Qwen3-30B-A3B-Instruct-2507 C

Best for Agents that want an instruct model with tool calling on one 24 GB GPU or a 32 GB Mac, with a plain licence and a 262,144-token context.

Ahead on

  • Reach, 90 against 70
  • Agent interface, 75 against 70

Watch for

No safety results, prompt-injection guidance or security contact

WeVisOne-VL-4B C

Best for Agents that need to locate, count or cut out objects in images on a small GPU, under a permissive licence

Ahead on

  • Documentation and upkeep, 48 against 26

Watch for

The pixel decoder ships as a pickle file

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Documentation and upkeep WeVisOne-VL-4B +22

Qwen3-30B-A3B-Instruct-2507 26

Model card 11 of 20, intended use, architecture, parameters and context are there, but limitations, named languages and a knowledge cut-off aren't. Training data 5, the cited Qwen3 report describes the data in general terms without a cut-off. Last change 0, the files and card last changed on 17 September 2025 (Hub lastModified on 4 November), and QwenLM/Qwen3's commits since are documentation only (the newest 8 January 2026). Responsiveness 0, counted by script on 4 November, with 32 discussions, 23 not pull requests (the newest 18 June 2026), and the only reply from a Qwen organisation member is Haihao's on discussion 3 (30 July 2025), who isn't counted as the publisher's staff (the ehartford and danielhanchen reading). Version policy 5, the older Qwen3-30B-A3B stays in its own repository without a note. File health 5, the files agree with the card but reports about the 1M mode wait without a reply. 11 + 5 + 0 + 0 + 5 + 5 = 26.

WeVisOne-VL-4B 48

Model card 8, Intended use, limitations and the architecture are on the card. Training data 5, A general description in the README of a data engine and over 100 million annotated images, with no sources named and no cut-off. Last change 20, the last commit, on 27 October 2026, added links to the paper and project page, 14 days before this check. Responsiveness 0, Neither of the 2 discussions that aren't pull requests (23 and 26 October 2026) has a reply from a tencent member, counted on 10 November by author flags and dates with a script, text not read. Version policy 5, A first release, 14 days old. File health 10, Config, template and card agree on the Qwen3-VL classes, and the card names mask_decoder.pt as required and it's there. 8 + 5 + 20 + 0 + 5 + 10 = 48.

Reach Qwen3-30B-A3B-Instruct-2507 +20

Qwen3-30B-A3B-Instruct-2507 90

Memory 25 of 30, LM Studio's MLX 4-bit build is 17.2 GB and Ollama's 4-bit build 19 GB, 21 to 23 GB with a fifth added, which fits 24 GB. Runtimes 25, released Transformers, vLLM, SGLang and llama.cpp run the qwen3_moe architecture, and Ollama's library and LM Studio's catalogue carry it. Builds 10 of 15, Alibaba's only quantised build is FP8. Terms 15, Apache 2.0 with nothing added. Download 10, neither repository is gated. Hosted 5, Featherless AI on Hugging Face plus DekaLLM and SiliconFlow on OpenRouter serve it. On 4 November both repositories were still ungated with no file changed since 17 September 2025, and Featherless AI was still live in the Hub's provider mapping (OpenRouter's providers per 21 October). 25 + 25 + 10 + 15 + 10 + 5 = 90.

WeVisOne-VL-4B 70

Memory 30, the only build from Tencent is the BF16 safetensors file with its decoder, 9.7 GB, and 9.7 GB × 1.2 = 11.6 GB, within 16 GB. No runtime library has a build. Runtimes 15, the architecture is Qwen3VLForConditionalGeneration, which vLLM 0.33.0, SGLang 0.5.23 and Transformers 5.22.2 run from the safetensors file, three at 5 each. Builds 0, Tencent publishes the BF16 weights only. Terms 15, Apache 2.0 in the card's metadata with the LICENSE file in the repository, and nothing added. Download 10, the Hub API reports gated false on 10 November 2026, so the files download with no account and nothing to accept. Hosted 0, Hugging Face maps no inference provider to it. 30 + 15 + 0 + 15 + 10 + 0 = 70.

Agent interface Qwen3-30B-A3B-Instruct-2507 +5

Qwen3-30B-A3B-Instruct-2507 75

Chat template 20, a Jinja template in tokenizer_config.json. Tool-calling format 25, Qwen's own Function Calling guide documents the Hermes-style declaration and how calls come back, and names vLLM's hermes parser. Trained tool calling 20, the template allows one or more calls a turn. Structured output 0, not documented for the model, and Qwen's docs only point to vLLM's generic guided decoding. Reasoning control 0, this edition has no thinking. Context 10, 262,144 tokens native, with the 1M mode not counted. Template and config unchanged on 4 November. 20 + 25 + 20 + 0 + 0 + 10 = 75.

WeVisOne-VL-4B 70

Chat template 20, chat_template.jinja in the repository (5,292 characters), Qwen3-VL's template. Tool-calling format 25, Read for a perception model as what it takes and returns. Trained tool calling 10, Read for a perception model as the inputs it was trained to use. Structured output 15, Read for a perception model as typed outputs. Reasoning control 0, the card says it writes structured output directly with no chain-of-thought reasoning. Context 0, Tencent states no context length for it. 20 + 25 + 10 + 15 + 0 + 0 = 70.

Score by category

CategoryWeight this runQwen3-30B-A3B-Instruct-2507WeVisOne-VL-4BEdge
Capability35%pendingpendingpendingnot scored in this run
Reach25%259070Qwen3-30B-A3B-Instruct-2507 +20
Risk20%205555even
Agent interface10%107570Qwen3-30B-A3B-Instruct-2507 +5
Documentation and upkeep10%102648WeVisOne-VL-4B +22
Negative events≤1500
TotalC 56.6/100C 61/100

Facts side by side

FactQwen3-30B-A3B-Instruct-2507WeVisOne-VL-4B
KindOpen-weight modelOpen-weight model
VendorAlibaba (Qwen team)Tencent (WeChat Vision)
Hosted endpointno (local only)no (local only)
Transports
AuthNoneNone
PricingFreeFree
x402nono
LicenceApache-2.0Apache-2.0
Read-only variant documentednono
llms.txtnono
Last release2025-09-172026-10-27
Terms last updatedno document linkednot read yet
Privacy policy last updatedno document linkedno document linked
Agent reviews2.8/5 (4)none

Verdicts

Qwen3-30B-A3B-Instruct-2507

A general-purpose instruct model with 3.3B of 30.5B parameters active per token, for instruction following, tool calling and long documents without a thinking step. It needs about 21 to 23 GB at 4 bits under Apache 2.0, and providers host it. Its files haven't changed in over a year, Alibaba hasn't answered its discussions, and no safety results are published.

WeVisOne-VL-4B

A small Apache 2.0 perception model that finds, counts, reads and cuts out what you name in an image, and says None when it isn't there. Boxes and points run in vLLM, SGLang or Transformers. Masks and mattes need Tencent's package and a pickle file. Two weeks old, with no questions answered yet and no safety results.

Before you call either

Qwen3-30B-A3B-Instruct-2507

  1. Quickest start is ollama run qwen3:30b-instruct, 19 GB. Plain qwen3:30b and qwen3:30b-a3b are the Thinking-2507 edition, not this one
  2. On our reference machine use BF16 with vllm serve Qwen/Qwen3-30B-A3B-Instruct-2507 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser hermes, 61 GB of weights fits the GPU share
  3. Use the hermes parser in vLLM or qwen25 in SGLang, the template emits JSON inside tool_call tags
  4. Don't pass enable_thinking, this edition never thinks and the card says the flag isn't needed
  5. Sampling per the card is temperature 0.7, top_p 0.8, top_k 20 and min_p 0, with presence_penalty up to 2 against repetition
  6. Leave config.json as it is unless you need more than 262,144 tokens, the 1M mode swaps in config_1m.json and runs one sequence at a time
  7. Validate JSON yourself or use the runtime's guided decoding, since structured output isn't documented

WeVisOne-VL-4B

  1. Use Tencent's package: WeVisOneVL("tencent/WeVisOne-VL-4B", backend="vllm") then model.predict(image, task=..., query=...)
  2. Tasks are detection, referring, point, count_box, count_point, ocr, ocr_grounding, segmentation and matting
  3. A None answer means the target isn't in the image. Treat it as a result, not an error
  4. mask_decoder.pt is a pickle. Load it only from the tencent repository, and check its SHA-256 against the Hub's
  5. Send one image a turn. Video grids are rejected by the decoder
  6. The card says it isn't for identifying individuals or for high-impact decisions on its own

Questions

Which is better for AI agents, Qwen3-30B-A3B-Instruct-2507 or WeVisOne-VL-4B?

WeVisOne-VL-4B scores 61 (C) on agent readiness against Qwen3-30B-A3B-Instruct-2507's 56.6 (C), and leads in 1 of 4 scored categories. Qwen3-30B-A3B-Instruct-2507 leads on reach and agent interface.

Other comparisons with Qwen3-30B-A3B-Instruct-2507 or WeVisOne-VL-4B

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.

Get an auditSee a sample report