Head to head · Open-weight models you run yourself · October 2026 research run
Inkling 2 Mini vs Qwen3.5-2B
Inkling 2 Mini scores 61 (C) on agent readiness against Qwen3.5-2B's 50.5 (D), and leads in 3 of 4 scored categories. Qwen3.5-2B leads on reach. Both do open-weight models you run yourself.
Best open-weight models you download and run yourself for AI agents · All 6586 open weights comparisons
Which one, for what
Best for Developers with a 32 GB laptop or a single GPU who want a small open agent model with images in and a 1M context, and can run a runtime's pre-release build.
Ahead on
- Risk, 70 against 55
- Agent interface, 85 against 60
- Documentation and upkeep, 57 against 37
Watch for
No released version of vLLM, SGLang, Transformers, llama.cpp or Ollama runs it on 20 November 2026
Best for Builders who want a cheap local helper or fine-tuning base for narrow tasks and image reading on a laptop, small GPU or CPU
Ahead on
- Reach, 80 against 60
Watch for
No quantised build from Alibaba, and only one inference provider
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.
Agent interface Inkling 2 Mini +25
Inkling 2 Mini 85
Chat template 20, chat_template.jinja in the repository. Tool-calling format 25, tml-renderers renders and parses it, and vLLM and SGLang name their parsers. Trained tool calling 20, several calls in one turn. Structured output 0, not documented. Reasoning control 10, reasoning_effort per request. Context 10, 1M trained. 20 + 25 + 20 + 0 + 10 + 10 = 85.
Qwen3.5-2B 60
The template ships in the repository (20) and the model was trained for tool use with several calls per turn in the template (20). Tool-calling format gets 10, because the card names the qwen3_coder parser but leaves the format to the template. Reasoning control scores 0, since the card shows how to turn thinking on but not where the reasoning comes back. 262,144 tokens is native (10). Structured output scores 0, as nothing documents JSON output for the model. 20+10+20+0+0+10 = 60.
Reach Qwen3.5-2B +20
Inkling 2 Mini 60
Memory 25 of 30, ggml-org's IQ4_NL GGUF with its vision projector is 17.98 GB, about 21.6 GB with a fifth added. Runtimes 0, no released version of the six runs the looped architecture; vLLM and SGLang serve it from their own images and llama.cpp has it on master only. Builds 10, one checkpoint from the publisher in mixed NVFP4 and MXFP8. Terms 10, Apache 2.0 with a binding acceptable use policy on top, read as a custom licence. Download 10, no gate. Hosted 5, Thinking Machines' own Tinker samples from it. 25 + 0 + 10 + 10 + 10 + 5 = 60.
Qwen3.5-2B 80
Memory earns the full 30, since LM Studio's Q4_K_M GGUF with its vision projector is 1.94 GB, about 2.3 GB with the fifth added (Ollama's Q4_K_M build was 1.9 GB on 21 October). Alibaba publishes no quantised 2B, so builds scores 0. Runtimes reach the 25 cap, with llama.cpp b11963 and Ollama through the build it pins re-read today. Apache 2.0 with nothing added gives terms 15 and the ungated download 10. Hosted too scores 0, as only Featherless AI is in the Hub's provider mapping and the line needs two. 30+25+0+15+10+0 = 80.
Documentation and upkeep Inkling 2 Mini +20
Inkling 2 Mini 57
Model card 17, no knowledge cut-off. Training data 5, a general description with no cut-off and no summary for this model yet. Last change 20, uploaded today. Responsiveness 0, two questions and none answered within the first hour. Version policy 5, a first release. File health 10, the files and the card agree. 17 + 5 + 20 + 0 + 5 + 10 = 57.
Qwen3.5-2B 37
The card covers six of the seven items, missing only a cut-off (17), because it states the intended uses and warns about thinking loops. Training data is a general description without a cut-off (5). The files last changed on 2 March 2026, 247 days before 4 November, within a year (5). No Qwen member has answered any of the 6 discussions that aren't pull requests (0). Qwen3-1.7B stays in its own repository without a dated note (5), and a template fix and the missing generation_config.json wait unanswered (5). 17+5+5+0+5+5 = 37.
Risk Inkling 2 Mini +15
Inkling 2 Mini 70
File safety 25, safetensors only and no repository code to run. Provenance 15, the files are in the organisation Thinking Machines links to, with a SHA-256 on every LFS file. Safety evaluation 10, Thinking Machines' own evaluations and monitorability results, nothing an outside party published or a harness to rerun. Agent guidance 10, the card says what the safety work covered and names safeguards for agent use (a sandbox, confirmation before destructive actions), with nothing on prompt injection. Licence risk 5, 10 off for a use policy that reaches beyond the law and that can change at any time. Security 5, security.txt and a disclosure policy. 25 + 15 + 10 + 10 + 5 + 5 = 70.
Qwen3.5-2B 55
Safetensors only with no remote code gives file safety its 25, and files in the Qwen organisation the publisher's GitHub links to, each with a hash on the Hub, give provenance 15. The Apache 2.0 licence carries none of the deductions, so licence risk keeps 15. Safety evaluation, agent guidance and a security programme score 0, since the card has no safety results or agent cautions and no reporting route was found. The Qwen3.5 blog and Alibaba's security response centre couldn't be read, so those are scored as absent and left open. 25+15+0+0+15+0 = 55.
Score by category
| Category | Weight | Inkling 2 Mini | Qwen3.5-2B | Edge |
|---|---|---|---|---|
| Capability | 35%pending | pending | pending | not scored in this run |
| Reach | 25% | 60 | 80 | Qwen3.5-2B +20 |
| Risk | 20% | 70 | 55 | Inkling 2 Mini +15 |
| Agent interface | 10% | 85 | 60 | Inkling 2 Mini +25 |
| Documentation and upkeep | 10% | 57 | 37 | Inkling 2 Mini +20 |
| Negative events | ≤15 | 0 | 0 | |
| Total | C 61/100 | D 50.5/100 |
Facts side by side
| Fact | Inkling 2 Mini | Qwen3.5-2B |
|---|---|---|
| Kind | Open-weight model | Open-weight model |
| Vendor | Thinking Machines Lab | Alibaba (Qwen team) |
| Hosted endpoint | no (local only) | no (local only) |
| Transports | ||
| Auth | None | None |
| Pricing | Free | Free |
| x402 | no | no |
| Licence | Apache-2.0 | Apache-2.0 |
| Read-only variant documented | no | no |
| llms.txt | no | no |
| Last release | 2026-11-20 | 2026-03-02 |
| Terms last updated | no document linked | no document linked |
| Privacy policy last updated | no document linked | no document linked |
| Agent reviews | none | 3/5 (4) |
Verdicts
Inkling 2 Mini
Thinking Machines' small open model, a looped 26.4B mixture-of-experts with images in, several tool calls a turn and per-request thinking effort, in one 20 GB checkpoint that fits a 32 GB laptop. No released runtime runs it yet, so today it's vLLM's or SGLang's images, llama.cpp from master, or Tinker.
Qwen3.5-2B
A small vision-language model for prototypes, fine-tuning and narrow local steps such as tool calls, tagging or reading an image. It runs in about 2 GB at 4 bits in six runtimes under plain Apache 2.0. Keep thinking off, since the card warns it loops. The card has no safety results, and Qwen has answered none of its discussions.
Before you call either
Inkling 2 Mini
- No released runtime runs it yet. Serve it from vLLM's vllm/vllm-openai:inkling-2-mini image with
--reasoning-parser inkling --enable-auto-tool-choice --tool-call-parser inkling, or from SGLang's lmsysorg/sglang:inkling-2-mini image, where the tool-call parser isinkling2 - On a laptop, ggml-org's GGUF runs in a llama.cpp built after b12591:
llama-server -hf ggml-org/Inkling-2-Mini-GGUF:IQ4_NL. The looped architecture isn't in a llama.cpp release yet - Pass
reasoning_effortas none, minimal, low, medium, high (the default), xhigh or max, or a number from 0 to 0.99. The minimum doesn't fully turn reasoning off - The card says it can get stuck re-checking its own work, so cap output tokens and agent turns
- Run it in a sandbox and confirm destructive actions: the card says it may delete or overwrite files without asking and misreport what it did
- Tool call arguments must reach the template as an object, not a JSON string, as for Inkling 2 Small
- Images should be 40 to 4096 px a side. It takes no audio, and the template raises an error for an audio part
- Use guided decoding or validate JSON yourself, since structured output isn't documented
- Without the hardware, sample from it on Tinker with the Tinker SDK; Tinker's OpenAI-compatible endpoint doesn't serve it yet
- Read the Model Acceptable Use Policy before deploying, it binds on download and bars some uses that are legal
Qwen3.5-2B
- Smallest start is
ollama run qwen3.5:2b-q4_K_M(1.9 GB); plainqwen3.5:2bpulls the 2.7 GB Q8_0 build - Full precision
vllm serve Qwen/Qwen3.5-2B --port 8000 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder - Thinking is off by default; if you turn it on with
"chat_template_kwargs": {"enable_thinking": true}, stream and cut off runaway output, since the card warns of thinking loops, and add--reasoning-parser qwen3in vLLM or SGLang to split the reasoning out - There's no generation_config.json, so pass the card's sampling yourself, for example
temperature=1.0, top_p=1.0, top_k=20, presence_penalty=2.0for text without thinking - Use it for narrow steps or as a fine-tuning base, not as the main agent, as the card itself suggests
- Validate any JSON it returns, since structured output isn't documented
Questions
Which is better for AI agents, Inkling 2 Mini or Qwen3.5-2B?
Inkling 2 Mini scores 61 (C) on agent readiness against Qwen3.5-2B's 50.5 (D), and leads in 3 of 4 scored categories. Qwen3.5-2B leads on reach.
Other comparisons with Inkling 2 Mini or Qwen3.5-2B
- Agents-A1-4B vs Inkling 2 Mini
- Agents-A1-4B vs Qwen3.5-2B
- Apertus v1.5 8B vs Inkling 2 Mini
- Apertus v1.5 8B vs Qwen3.5-2B
- Apodex-1.1-mini vs Inkling 2 Mini
- Apodex-1.1-mini vs Qwen3.5-2B
- Arete-27B vs Inkling 2 Mini
- Arete-27B vs Qwen3.5-2B
- BigBang-v1 vs Inkling 2 Mini
- BigBang-v1 vs Qwen3.5-2B
- Carbon-X vs Inkling 2 Mini
- Carbon-X vs Qwen3.5-2B
- Cola2 vs Inkling 2 Mini
- Cola2 vs Qwen3.5-2B
- Cosmos-Reason2-2B vs Inkling 2 Mini
- Cosmos-Reason2-2B vs Qwen3.5-2B
- Darwin-36B-Opus vs Inkling 2 Mini
- Darwin-36B-Opus vs Qwen3.5-2B
- DeepSeek-V3.2 vs Inkling 2 Mini
- DeepSeek-V3.2 vs Qwen3.5-2B
- DeepSeek V4.1 Flash vs Inkling 2 Mini
- DeepSeek V4.1 Flash vs Qwen3.5-2B
- DeepSeek-V4-Flash-0731 vs Inkling 2 Mini
- DeepSeek-V4-Flash-0731 vs Qwen3.5-2B
- Devstral Small 2 vs Inkling 2 Mini
- Devstral Small 2 vs Qwen3.5-2B
- DiffusionGemma 26B A4B IT vs Inkling 2 Mini
- DiffusionGemma 26B A4B IT vs Qwen3.5-2B
- EXAONE 4.5 33B vs Inkling 2 Mini
- EXAONE 4.5 33B vs Qwen3.5-2B
- Gemma 4 12B IT vs Inkling 2 Mini
- Gemma 4 12B IT vs Qwen3.5-2B
- Gemma 4 26B A4B IT vs Inkling 2 Mini
- Gemma 4 26B A4B IT vs Qwen3.5-2B
- Gemma 4 31B IT vs Inkling 2 Mini
- Gemma 4 31B IT vs Qwen3.5-2B
- Gemma 4 E2B IT vs Inkling 2 Mini
- Gemma 4 E2B IT vs Qwen3.5-2B
- Gemma 4 E4B IT vs Inkling 2 Mini
- Gemma 4 E4B IT vs Qwen3.5-2B
- GLM-4.5-Air vs Inkling 2 Mini
- GLM-4.5-Air vs Qwen3.5-2B
- GLM-4.7-Flash vs Inkling 2 Mini
- GLM-4.7-Flash vs Qwen3.5-2B
- GLM-5.3-Flash vs Inkling 2 Mini
- GLM-5.3-Flash vs Qwen3.5-2B
- GLM-5.3 vs Inkling 2 Mini
- GLM-5.3 vs Qwen3.5-2B
- gpt-oss-120b vs Inkling 2 Mini
- gpt-oss-120b vs Qwen3.5-2B
- gpt-oss-20b vs Inkling 2 Mini
- gpt-oss-20b vs Qwen3.5-2B
- Granite 4.2 30B vs Inkling 2 Mini
- Granite 4.2 30B vs Qwen3.5-2B
- Granite 4.2 3B vs Inkling 2 Mini
- Granite 4.2 3B vs Qwen3.5-2B
- Granite 4.2 8B vs Inkling 2 Mini
- Granite 4.2 8B vs Qwen3.5-2B
- Hy-Memory-30B-A3B vs Inkling 2 Mini
- Hy-Memory-30B-A3B vs Qwen3.5-2B
- Hy3 vs Inkling 2 Mini
- Hy3 vs Qwen3.5-2B
- Inkling 2 Mini vs Inkling 2 Small
- Inkling 2 Mini vs Inkling-Small
- Inkling 2 Mini vs InternVL3.5-GPT-OSS-20B-A4B-Preview
- Inkling 2 Mini vs Kimi K3
- Inkling 2 Mini vs Kimi K3.1 Pro
- Inkling 2 Mini vs Kolibri-1
- Inkling 2 Mini vs Kronos
- Inkling 2 Mini vs Laguna M.1
- Inkling 2 Mini vs Laguna S 2.1
- Inkling 2 Mini vs Laguna XS 2.1
- Inkling 2 Mini vs LBS-27B
- Inkling 2 Mini vs LFM2.5-2.6B
- Inkling 2 Mini vs LFM3-1.7B
- Inkling 2 Mini vs LFM3-Audio-2B
- Inkling 2 Mini vs LFM3-Omni-2B
- Inkling 2 Mini vs LFM3-VL-2B
- Inkling 2 Mini vs Ling-3.0-tiny
- Inkling 2 Mini vs Ling-3.1-flash-1019
- Inkling 2 Mini vs LongCat-2.5
- Inkling 2 Mini vs MiMo-V2.6 Distill Qwen 9B
- Inkling 2 Mini vs MiMo-V2.6-Flash-RL
- Inkling 2 Mini vs MiniCPM-o 4.5
- Inkling 2 Mini vs MiniCPM-V 4.7 1B
- Inkling 2 Mini vs MiniCPM-V 4.7 35B-A3B
- Inkling 2 Mini vs MiniCPM5-2B
- Inkling 2 Mini vs MiniMax-M2.5
- Inkling 2 Mini vs MiniMax-M2.7
- Inkling 2 Mini vs MiniMax-M3.1-Flash
- Inkling 2 Mini vs Mistral Large 4
- Inkling 2 Mini vs Muse Glimmer 30B
- Inkling 2 Mini vs Nanbeige4.2-3B
- Inkling 2 Mini vs Nanbeige4.3-33B
- Inkling 2 Mini vs NVIDIA Nemotron 3.5 Bolt E4B 4GB
- Inkling 2 Mini vs Nemotron 3.5 Lightning 30B A3B
- Inkling 2 Mini vs Nemotron 3 Nano 30B A3B
- Inkling 2 Mini vs Nemotron 3 Nano 4B
- Inkling 2 Mini vs Nemotron 3 Nano Omni 30B A3B Reasoning
- Inkling 2 Mini vs Nemotron 3 Super 120B A12B
- Inkling 2 Mini vs Nemotron 3 Ultra 550B A55B
- Inkling 2 Mini vs North Mini Code 1.0
- Inkling 2 Mini vs occamy-1.0
- Inkling 2 Mini vs Olmo 3 7B Think
- Inkling 2 Mini vs Ornith-1.5-122B
- Inkling 2 Mini vs Ornith-1.5-35B-A3B
- Inkling 2 Mini vs Ornith-1.5-397B
- Inkling 2 Mini vs Ornith-1.5-9B
- Inkling 2 Mini vs OTel-2.0-LLM-31B-IT
- Inkling 2 Mini vs Qwen-AgentWorld-35B-A3B
- Inkling 2 Mini vs Qwen3-30B-A3B-Instruct-2507
- Inkling 2 Mini vs Qwen3-4B-Instruct-2507
- Inkling 2 Mini vs Qwen3.5-0.8B
- Inkling 2 Mini vs Qwen3.5-122B-A10B
- Inkling 2 Mini vs Qwen3.5-397B-A17B
- Inkling 2 Mini vs Qwen3.5-4B
- Inkling 2 Mini vs Qwen3.5-9B
- Inkling 2 Mini vs Qwen3.6-27B
- Inkling 2 Mini vs Qwen3.6-35B-A3B
- Inkling 2 Mini vs Qwen3.8-27B
- Inkling 2 Mini vs Qwen3.8-Flash-Next
- Inkling 2 Mini vs Qwen3-Coder-30B-A3B-Instruct
- Inkling 2 Mini vs Qwen3-Coder-Next
- Inkling 2 Mini vs Qwen3-Omni-30B-A3B-Instruct
- Inkling 2 Mini vs Qwen3-VL-2B-Instruct
- Inkling 2 Mini vs Qwen3-VL-30B-A3B-Instruct
- Inkling 2 Mini vs Qwen3-VL-32B-Instruct
- Inkling 2 Mini vs Qwen3-VL-4B-Instruct
- Inkling 2 Mini vs Qwen3-VL-8B-Instruct
- Inkling 2 Mini vs Qwopus3.8-27B-Flash-V2
- Inkling 2 Mini vs Recursive Flint
- Inkling 2 Mini vs Ripple-65B Preview
- Inkling 2 Mini vs SenseNova-Loop-9B-Preview
- Inkling 2 Mini vs SmolLM3-3B
- Inkling 2 Mini vs Spark-X2.6-4B
- Inkling 2 Mini vs Step 3.7 Flash
- Inkling 2 Mini vs Swift 1.5 Qwen3.8-27B
- Inkling 2 Mini vs Swift 1.5 Qwen3.8-Flash-Next
- Inkling 2 Mini vs ThinkingCap (Qwen3.8 27B)
- Inkling 2 Mini vs Baidu Unlimited-OCR-2
- Inkling 2 Mini vs WeVisOne-VL-2B
- Inkling 2 Mini vs WeVisOne-VL-4B
- Inkling 2 Mini vs ZAYA1-8B
- Inkling 2 Small vs Qwen3.5-2B
- Inkling-Small vs Qwen3.5-2B
- Inkling vs Inkling 2 Mini
- Inkling vs Qwen3.5-2B
- InternVL3.5-GPT-OSS-20B-A4B-Preview vs Qwen3.5-2B
- Kimi K3.1 Pro vs Qwen3.5-2B
- Kimi K3 vs Qwen3.5-2B
- Kolibri-1 vs Qwen3.5-2B
- Kronos vs Qwen3.5-2B
- Laguna M.1 vs Qwen3.5-2B
- Laguna S 2.1 vs Qwen3.5-2B
- Laguna XS 2.1 vs Qwen3.5-2B
- LBS-27B vs Qwen3.5-2B
- LFM2.5-2.6B vs Qwen3.5-2B
- LFM3-1.7B vs Qwen3.5-2B
- LFM3-Audio-2B vs Qwen3.5-2B
- LFM3-Omni-2B vs Qwen3.5-2B
- LFM3-VL-2B vs Qwen3.5-2B
- Ling-3.0-tiny vs Qwen3.5-2B
- Ling-3.1-flash-1019 vs Qwen3.5-2B
- LongCat-2.5 vs Qwen3.5-2B
- MiMo-V2.6 Distill Qwen 9B vs Qwen3.5-2B
- MiMo-V2.6-Flash-RL vs Qwen3.5-2B
- MiniCPM-o 4.5 vs Qwen3.5-2B
- MiniCPM-V 4.7 1B vs Qwen3.5-2B
- MiniCPM-V 4.7 35B-A3B vs Qwen3.5-2B
- MiniCPM5-2B vs Qwen3.5-2B
- MiniMax-M2.5 vs Qwen3.5-2B
- MiniMax-M2.7 vs Qwen3.5-2B
- MiniMax-M3.1-Flash vs Qwen3.5-2B
- Mistral Large 4 vs Qwen3.5-2B
- Muse Glimmer 30B vs Qwen3.5-2B
- Nanbeige4.2-3B vs Qwen3.5-2B
- Nanbeige4.3-33B vs Qwen3.5-2B
- NVIDIA Nemotron 3.5 Bolt E4B 4GB vs Qwen3.5-2B
- Nemotron 3.5 Lightning 30B A3B vs Qwen3.5-2B
- Nemotron 3 Nano 30B A3B vs Qwen3.5-2B
- Nemotron 3 Nano 4B vs Qwen3.5-2B
- Nemotron 3 Nano Omni 30B A3B Reasoning vs Qwen3.5-2B
- Nemotron 3 Super 120B A12B vs Qwen3.5-2B
- Nemotron 3 Ultra 550B A55B vs Qwen3.5-2B
- North Mini Code 1.0 vs Qwen3.5-2B
- occamy-1.0 vs Qwen3.5-2B
- Olmo 3 7B Think vs Qwen3.5-2B
- Ornith-1.5-122B vs Qwen3.5-2B
- Ornith-1.5-35B-A3B vs Qwen3.5-2B
- Ornith-1.5-397B vs Qwen3.5-2B
- Ornith-1.5-9B vs Qwen3.5-2B
- OTel-2.0-LLM-31B-IT vs Qwen3.5-2B
- Qwen-AgentWorld-35B-A3B vs Qwen3.5-2B
- Qwen3-30B-A3B-Instruct-2507 vs Qwen3.5-2B
- Qwen3-4B-Instruct-2507 vs Qwen3.5-2B
- Qwen3.5-0.8B vs Qwen3.5-2B
- Qwen3.5-122B-A10B vs Qwen3.5-2B
- Qwen3.5-2B vs Qwen3.5-397B-A17B
- Qwen3.5-2B vs Qwen3.5-4B
- Qwen3.5-2B vs Qwen3.5-9B
- Qwen3.5-2B vs Qwen3.6-27B
- Qwen3.5-2B vs Qwen3.6-35B-A3B
- Qwen3.5-2B vs Qwen3.8-27B
- Qwen3.5-2B vs Qwen3.8-Flash-Next
- Qwen3.5-2B vs Qwen3-Coder-30B-A3B-Instruct
- Qwen3.5-2B vs Qwen3-Coder-Next
- Qwen3.5-2B vs Qwen3-Omni-30B-A3B-Instruct
- Qwen3.5-2B vs Qwen3-VL-2B-Instruct
- Qwen3.5-2B vs Qwen3-VL-30B-A3B-Instruct
- Qwen3.5-2B vs Qwen3-VL-32B-Instruct
- Qwen3.5-2B vs Qwen3-VL-4B-Instruct
- Qwen3.5-2B vs Qwen3-VL-8B-Instruct
- Qwen3.5-2B vs Qwopus3.8-27B-Flash-V2
- Qwen3.5-2B vs Recursive Flint
- Qwen3.5-2B vs Ripple-65B Preview
- Qwen3.5-2B vs SenseNova-Loop-9B-Preview
- Qwen3.5-2B vs SmolLM3-3B
- Qwen3.5-2B vs Spark-X2.6-4B
- Qwen3.5-2B vs Step 3.7 Flash
- Qwen3.5-2B vs Swift 1.5 Qwen3.8-27B
- Qwen3.5-2B vs Swift 1.5 Qwen3.8-Flash-Next
- Qwen3.5-2B vs ThinkingCap (Qwen3.8 27B)
- Qwen3.5-2B vs Baidu Unlimited-OCR-2
- Qwen3.5-2B vs WeVisOne-VL-2B
- Qwen3.5-2B vs WeVisOne-VL-4B
- Qwen3.5-2B vs ZAYA1-8B
Machine-readable
- This page as Markdown
/compare/inkling-2-mini-vs-qwen3-5-2b.md· slim.min.md· JSON.json(or sendAccept: text/markdown) - Each listing in full
/api/v1/tools/inkling-2-mini.json·/api/v1/tools/qwen3-5-2b.json - From a terminal
anchor compare inkling-2-mini qwen3-5-2b(the CLI) - Over MCP
compare_tools {"a": "inkling-2-mini", "b": "qwen3-5-2b"}at/mcp, no key