DeepSeek V4.1 Flash vs Step 3.7 Flash

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.

Step 3.7 Flash 70

Memory 5, the smallest build at 4 bits is StepFun's IQ4_XS GGUF, 104.99 GB with the 3.97 GB projector, and 108.96 GB × 1.2 = 130.8 GB, over what the reference machine can give a model and within 256 GB. The IQ3 and Q3 builds are below 4 bits and aren't read. Runtimes 20, Released vLLM 0.33.0, SGLang 0.5.23 and Transformers 5.22.2 have the step3p7 model, and llama.cpp b12242 has the architecture, with StepFun's own GGUF. Four at 5 each, and the card documents all four. Builds 15, StepFun publishes GGUF at five quantised precisions, an FP8 build and an NVFP4 build. Terms 15, Apache 2.0 in the card's metadata and its licence section, an OSI licence with nothing added. Download 10, the Hub API reports gated false for all four repositories on 10 November 2026, so the files download with no account and nothing to accept. Hosted 5, StepFun's own API serves it, per the card and our StepFun API listing, and Featherless and DeepInfra are live for it on Hugging Face. 5 + 20 + 15 + 15 + 10 + 5 = 70.

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).

Step 3.7 Flash 85

Chat template 20, chat_template.jinja in the repository (5,723 characters). Tool-calling format 25, the card names the parser to use in two runtimes, step3p5 in vLLM (with auto tool choice) and in SGLang, and both ship Step parsers in a release. Trained tool calling 20, the card presents tool orchestration as a main use, with multi-turn results, and the template renders every call in a turn. Structured output 0, Nothing on the card about JSON or schema-constrained output for the weights. Reasoning control 10, the template has a reasoning_effort setting and keeps reasoning_content, and the card names the step3p5 reasoning parser for vLLM and SGLang, which returns the reasoning apart from the answer. Context 10, 256K tokens per the card, and 262,144 positions in config.json. 20 + 25 + 20 + 0 + 10 + 10 = 85.

DeepSeek V4.1 Flash 48

The cards cover architecture, parameters and context but not intended use, limitations, languages or a cut-off, so 8. Training data is a general description with no cut-off, which caps it at 5. The files changed on 16 and 17 October and DeepSeek's deepseek-recipe 0.2.0 changed the V4.1 prompt encoding on 20 October, inside 30 days (20). None of the 40 most recent questions across both repositories had a DeepSeek reply (0). Every release keeps its own repository and the update names what it replaces (10). File health loses 5 for an open, unanswered pull request saying the September template trails the encoding changes. 8 + 5 + 20 + 0 + 10 + 5 = 48.

Step 3.7 Flash 39

Model card 14, Intended use, the architecture, the parameters, the context and the language (in the metadata) are on the card. Training data 0, Nothing on the card about the training data. Last change 10, the last commit, on 3 June 2026, fixed the modelling code, 160 days before this check and within 180. Responsiveness 0, none of the 11 most recent discussions that aren't pull requests (29 May to 24 June 2026) has a reply from a stepfun-ai member, counted on 10 November by author flags and dates with a script, text not read. Version policy 5, Step-3.5-Flash stays in its own repositories and 3.7 came out in new ones, with no dated note on the older cards that we read. File health 10, Config, template and card agree on the architecture and the context, and June commits fixed the modelling code. 14 + 0 + 10 + 0 + 5 + 10 = 39.

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.

Step 3.7 Flash 45

File safety 15, safetensors and GGUF only, no .bin, .pt or .pth files (25). Provenance 15, the files sit in the stepfun-ai organisation, which the card's StepFun platform and model page belong to (10). Safety evaluation 0, no safety results on the card. Agent guidance 0, the card has no limitations section and gives no guidance on prompt injection, instruction priority, refusals or safeguards. Licence risk 15, Apache 2.0 has no use policy, threshold or change clause. Security programme 0, no way to report a flaw on the card, and our StepFun API listing found no security.txt, bug bounty or disclosure policy. 15 + 15 + 0 + 0 + 15 + 0 = 45.