Open Weight Thoughts

Open Weight Thoughts

The latest in open source AI & real opinions of software engineers

  1. Safety Testing That Stops at the Download Button

    The reported White House framework would voluntarily test powerful closed AI models for security risks while excluding open-weight releases. That may protect open development from a bad regulatory fit, but it also treats model distribution as if it changes what a model can do.

    · Q. Wang· 7 min read· guides· humor

  2. Qwen3.7-Max Non-Quantized API & Coding-Agent Support

    Qwen3.7-Max is available through QwenCloud’s hosted API, but “non-quantized API” is not a meaningful product option: Alibaba does not expose downloadable Qwen3.7-Max weights or the serving precision behind the endpoint. Coding agents that accept an OpenAI-compatible endpoint, including Cline and Qwen Code with custom provider settings, can be configured to call the exact `qwen3.7-max` model ID.

    · Q. García· 7 min read· guides

  3. Why High Benchmark Scores Still Break Your JSON

    AI benchmarks increasingly resemble children’s sports leagues: nearly every serious model gets a medal, while your production parser quietly discovers that “valid JSON only” was apparently an artistic suggestion. Here is what saturated scores mean, why format following is harder than it looks, and how engineers should evaluate models instead.

    · S. Yang· 7 min read· guides· humor

  4. Why OpenAI Can Promote Open Weights and Still Keep the Best Models Behind an API

    OpenAI’s open-weight strategy is not a retreat from closed models. It is a sensible two-track product plan: sell the frontier as a managed service, and let developers run capable older or narrower models wherever they need control.

    · V. Martínez· 7 min read· guides· humor

  5. MiMo-V2.5 Non-Quantized API for OpenRouter, Cursor & Cline

    MiMo-V2.5 is available through OpenRouter and Xiaomi’s official API, but “non-quantized” is not an accurate promise for the released model or a routed endpoint. Learn what the published FP8 weights mean, how to choose between the official provider and OpenRouter, and why Cline is a safer coding-agent integration than Cursor.

    · T. Shevchenko· 8 min read· guides

  6. MiMo-V2.5-Pro on OpenRouter: Claude Code, OpenCode & Kilo

    MiMo-V2.5-Pro is available on OpenRouter as `xiaomi/mimo-v2.5-pro` and can be used with Claude Code, OpenCode, and Kilo Code. The important caveat is that “non-quantized” is not a provider choice currently advertised for this model: Xiaomi’s official downloadable release is FP8 mixed precision, while OpenRouter exposes provider-level precision filtering only when matching endpoints exist.

    · I. Liu· 8 min read· guides

  7. Local AI Should Be the Default Boundary, Not a Hobbyist Flex

    Local AI is not mainly about avoiding API bills or proving you own a GPU. It is a better default architecture for software teams because it makes data movement, failure modes, and model dependence explicit.

    · T. Novák· 8 min read· opinion· guides

  8. Qwen3.7-Max on OpenRouter: Cline, Roo Code, Non-Quantized

    Qwen3.7-Max is available through OpenRouter for cloud-based use in Cline and Roo Code, but there is no official downloadable non-quantized Qwen3.7-Max checkpoint for local inference. For coding agents, that means you can use the model behind one API key today, while treating it as a hosted model rather than an open-weight model you can run or fine-tune yourself.

    · B. Yamamoto· 7 min read· guides
Open Weight Thoughts