Reach for this when a bigger model does the thinking and you need the answer to validate against a schema every time. It runs locally, costs nothing, and ships as safetensors or GGUF. Do not expect general knowledge, and do not run it without an engine that enforces the schema.
Osmosis-Structure-0.6B is an open-weight small language model built on Qwen3 with 0.6 billion parameters. It converts reasoning traces and natural language reports into schema-valid JSON, attending only to the value for each key that the inference engine declares, and is served locally through SGLang.
Provider: Osmosis · Family: Osmosis
Input modalities: text · Output: text
About Osmosis-Structure-0.6B
Osmosis-Structure-0.6B is a small language model released by Osmosis on 28 May 2025 under a permissive open licence. It is built on Qwen3-0.6B and carries the same 0.6 billion parameters, which places it firmly in the category of models that run on a laptop rather than a cluster. It exists to do one job: take a messy reasoning trace or a natural language report and emit output that matches a caller-supplied JSON schema. The design choice that makes it work is narrow scope. During training the model was forced to attend only to the value for each key declared by the inference engine, rather than generating the whole document structure itself. The schema is the inference engine's responsibility; the model only fills in what goes in each slot. That division is why a model this small can hold its own on a task where much larger general models produce plausible prose that fails validation. Training ran in two stages. A baseline format was established from ten samples of randomly generated text paired with their JSON interpretations, then reinforcement learning was applied over roughly 500,000 pairs of JSON and natural language, consisting of reasoning traces with their final outputs or natural language reports with their expected structured formats. Osmosis used verl as the training framework and SGLang as the rollout backend, and modified parts of verl so a per-sample schema could be passed into the training data. The evaluation tables published on the model card are widely misread, so it is worth being precise. They measure what the osmosis structured-generation technique does when applied to other models, not what this model scores. On the Math DAPO 17K dataset, Claude 4 Sonnet moves from 15.52% to 69.40% with osmosis enhancement, GPT-4.1 from 10.53% to 70.03%, and OpenAI o3 from 91.14% to 94.05%. On AIME 1983-2024 the same pattern holds, with GPT-4.1 moving from 2.79% to 39.66%. Osmosis describes this as the same technique that powers the 0.6B model; it does not publish standalone benchmark scores for the model itself. Access is straightforward because there is no API and no price. Weights ship in safetensors and GGUF, the licence puts no commercial condition in your way, and Osmosis recommends serving through SGLang, whose launch command and an OpenAI-compatible Python client example are both given on the card. Because the interface is OpenAI-compatible, dropping it into an existing pipeline is mostly a base URL change. Adoption sits at roughly 8,000 downloads and over 400 likes, which is respectable for a specialist model and modest in absolute terms. It suits a pipeline that already has a capable model doing the reasoning and needs the last mile to be reliable: extracting a final answer, normalising a report into records, or guaranteeing a tool call validates. It is the wrong choice as a general assistant, since at this size it has no broad world knowledge, and the wrong choice if your serving stack cannot enforce a schema, because the schema enforcement is where half the value lives. Teams already committed to constrained decoding in vLLM or Outlines may find they have most of this without adding a model. The governance picture is thin, which is normal for a seed-stage release and worth stating plainly. There is no system card, no published safety evaluation, no red-teaming disclosure and no stated training-data provenance beyond the synthetic pair description above. The model has not been updated since 13 June 2025. For a component that only reshapes text into JSON that matters less than it would for a conversational model, but it is not nothing if the traces you are reshaping contain regulated data.
Pricing
Free. The weights are openly licensed and there is no hosted endpoint or per-token rate, so the only cost is the hardware you serve it on. At 0.6B parameters that is a modest GPU or a CPU with GGUF quantization.
Frequently Asked Questions
How much does Osmosis-Structure-0.6B cost in 2026?
Nothing. The weights are released under Apache 2.0 with no hosted endpoint and no per-token rate, so your only expense is the machine you run it on. Because it is a 0.6 billion parameter model with a GGUF build available, that can be a small GPU or even a CPU.
Is Osmosis-Structure-0.6B free to use commercially?
Yes. Apache 2.0 permits commercial use, modification and redistribution, including in closed-source products, provided you keep the licence and attribution. There is no separate commercial tier to buy, unlike the gated Osmosis-Apply-1.7B in the same organisation.
What are the best alternatives to Osmosis-Structure-0.6B?
Constrained decoding libraries such as Outlines or the structured output modes in vLLM and SGLang solve much of the same problem without adding a model. A larger instruct model with a JSON mode is the simpler option if you already pay for one. Pick this model when you want the extraction step isolated and cheap.
How does it compare to using GPT-4.1 with JSON mode in 2026?
They are not really rivals: the model card's own tables show the osmosis technique lifting GPT-4.1 from 10.53% to 70.03% on Math DAPO 17K, meaning the method is meant to sit alongside a large model rather than replace it. Use a frontier model for the reasoning, and a small local model like this for the final structured extraction to cut cost and latency.
How do you get started with Osmosis-Structure-0.6B?
Serve it with SGLang using the launch command on the model card, pointing at osmosis-ai/Osmosis-Structure-0.6B, then call it through any OpenAI-compatible client by setting the base URL to your local server. Pass your JSON schema, in raw form or dumped from a Pydantic model, along with the reasoning trace you want structured.