Devstral Small 2 vs Qwen3.5-9B

Devstral Small 2 46

Model card 11, Intended use, the architecture, the parameters and the context are on the card. Training data 0, Nothing on the card about the training data. Last change 10, the last commit, on 15 July 2026, updated the files to match mistral-common, 118 days before this check and within 180. Responsiveness 10, Of the 23 most recent discussions that aren't pull requests (9 December 2025 to 6 November 2026), 11 have a reply from a mistralai member, which is some. Version policy 5, Devstral Small 1.1 stays in its own repository, which this card links as the predecessor, with no dated note on the older card that we read. File health 10, Config, template, tokenizer and card agree, and the June and July commits fixed stray tokens and brought the files in line with mistral-common. 11 + 0 + 10 + 10 + 5 + 10 = 46.

Qwen3.5-9B 34

Model card 14, it covers intended use, architecture, parameters, context and 201 languages but gives no limitations and no knowledge cut-off. Training data 5, a general line about trillions of multimodal tokens with no cut-off. Last change 5, the files and card last changed on 2 March 2026 (c202236), 247 days before 4 November, within a year. Responsiveness 0, no Qwen member replied in any of the 40 most recent of the 45 discussions that aren't pull requests (2 March to 29 September 2026). Version policy 5, a first release that was never replaced in place, with no dated notes across generations. File health 5, two reports about the files and card (15 and 51) are still open with no reply. 14 + 5 + 5 + 0 + 5 + 5 = 34. The card's highlights mention sparse Mixture-of-Experts for the series while this 9B is dense, but its own overview matches config.json, so it's not counted as a contradiction.

Devstral Small 2 90

Memory 25, the smallest build at 4 bits from a runtime's own library is Ollama's devstral-small-2 at Q4_K_M, 15 GB, and 15 GB × 1.2 = 18 GB, within 24 GB. Mistral's own FP8 files are 25.79 GB. Runtimes 25, Released vLLM 0.33.0, SGLang 0.5.23 and Transformers 5.22.2 have the mistral3 model, llama.cpp b12242 has the architecture with a GGUF from its own organisation, and Ollama's library has devstral-small-2. Builds 10, one quantised precision from Mistral, the FP8 weights it releases. Terms 15, Apache 2.0 in the card's metadata and text, an OSI licence with 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 5, Mistral's own API serves Devstral Small 2: the card says the model is available through its API and points there when a local runtime underperforms. 25 + 25 + 10 + 15 + 10 + 5 = 90.

Qwen3.5-9B 85

Memory 30, Qwen publishes no quantised 9B, so the smallest 4-bit build from a runtime's own library is Ollama's qwen3.5:9b at 6.6 GB, about 7.9 GB with the fifth added, well under 16 GB. Runtimes 25, the cap, with vLLM listed in its stable docs, Transformers at released tags, llama.cpp, Ollama's library, LM Studio's own catalogue and SGLang all running the architecture. Builds 0, Qwen's FP8 and GPTQ repositories cover only the 27B and larger Qwen3.5 models, so every quantisation of the 9B is someone else's. Terms 15, plain Apache 2.0. Download 10, the repository isn't gated. Hosted 5, Together, Featherless AI, OVHcloud and DeepInfra are live for it on its Hugging Face page. 30 + 25 + 0 + 15 + 10 + 5 = 85.

Devstral Small 2 60

File safety 25, safetensors only, no .bin, .pt or .pth files and no Python in the repository. Provenance 15, files in Mistral's verified mistralai organisation, which Mistral's site links to per our Mistral Large 4 listing (10). Safety evaluation 0, no safety results on the card. Agent guidance 0, the card has no limitations or safety section, and nothing on prompt injection, instruction priority or what the model refuses. Licence risk 15, Apache 2.0 has no use policy, threshold or change clause. Security programme 5, mistral.ai's security.txt gives a security contact and a submission policy, as read for the Mistral Large 4 listing (5). 25 + 15 + 0 + 0 + 15 + 5 = 60.

Qwen3.5-9B 55

File safety 25, safetensors only and nothing from the repository runs on load or to build prompts. Provenance 15, the Qwen organisation linked from Alibaba's QwenLM README, with a SHA-256 for every shard. Safety evaluation 0, none on the card, and the Qwen3.5 blog couldn't be read, so this is couldn't check scored as absent. Agent guidance 0, the card has no prompt-injection guidance, refusal scope or cautions. Licence risk 15, Apache 2.0 with nothing added and a working licence link. Security programme 0, no SECURITY.md in Qwen's GitHub repositories, no security.txt on qwen.ai or alibabacloud.com, and whether Alibaba's group security centre takes reports about Qwen models couldn't be checked.

Devstral Small 2 65

Chat template 20, chat_template.jinja in the repository (11,362 characters), with the tokenizer's tekken file for mistral-common. Tool-calling format 25, the card documents tools in the OpenAI function format with an example, names the mistral tool-call parser with auto tool choice for vLLM, shows the same for SGLang, and ships test prompts with the expected tool-call output. Trained tool calling 10, the card says the model excels at using tools, and its tests cover one tool call and calls made one at a time in sequence. Structured output 0, Nothing on the card about JSON or schema-constrained answers for this model. Reasoning control 0, no reasoning mode is documented, and a June 2026 commit removed stray thinking tokens from the files. Context 10, 256K tokens per the card, served at 262,144 in its vLLM command. 20 + 25 + 10 + 0 + 0 + 10 = 65.

Qwen3.5-9B 70

Chat template 20, a Jinja file in the repository. Tool-calling format 10, the card names the qwen3_coder parser for vLLM and SGLang but leaves the declaration and return format to the template, and Qwen's own function-calling guide still describes Qwen3's Hermes format. Trained tool calling 20, the template renders several calls in one turn. Structured output 0, nothing documents asking the weights for JSON. Reasoning control 10, thinking turns off per request with enable_thinking false and the reasoning comes back in think tags that the qwen3 parser splits out. Context 10, 262,144 tokens native.