DeepSeek V4.1 Flash vs Devstral Small 2

DeepSeek V4.1 Flash 60

Memory scores nothing, because the only build is DeepSeek's own at 510.3 GB, about 612 GB with the fifth added, and nobody's runtime library holds a smaller one at 4-bit or better. Four runtimes run it in a released version (5 each), vLLM 0.33.0, SGLang 0.5.23, which maps the V4.1 architecture name onto its V4 model, llama.cpp from b11870 (25 October 2026), which added the deepseek41 architecture and its converter, and Ollama 0.40.14, which builds llama.cpp b11963. Transformers 5.22.1 and LM Studio don't (0 each). The release itself is one quantised build, FP8 with FP4 experts (10). MIT with nothing added earns the full terms line (15), the repositories aren't gated (10), and Hugging Face maps three live inference providers for the update (5). 0 + 20 + 10 + 15 + 10 + 5 = 60.

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.

DeepSeek V4.1 Flash 50

The weights are safetensors only, but 10 comes off because the card's way to build prompts is running encoding.py from the repository, since there's no Jinja template. The files sit in DeepSeek's own organisation with a hash for every file. Nothing is published on safety and there's no guidance for agent use, so those two lines score 0. The security programme earns 5 for the security@deepseek.com reporting address in DeepSeek's site footer, read as for DeepSeek-V3.2, with no security.txt and no public advisories. The MIT licence carries no use policy, threshold or revocation, so licence risk keeps all 15.

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.

DeepSeek V4.1 Flash 75

There's no Jinja template in the 17 October repository, only code to run, so the template line gets 10. DeepSeek documents the DSML tool format both ways and ships its own parser in encoding.py and deepseek-recipe, but names no parser in a runtime (10, read as for DeepSeek-V3.2 and V4-Flash-0731), and one block can carry several calls (20). deepseek-recipe documents JSON Schema output via response_format (15). Thinking turns off in chat mode and takes an effort budget from 1 to 100, with reasoning returned in think tags (10). The 1M context was trained in (10).

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.