Free Qwen 3.8 27B API: Not Yet — What to Run Instead

No — as of August 12, 2026 there is no free Qwen3.8-27B API, and there is no paid one either. The model was announced on August 3 with open weights promised to Hugging Face and ModelScope "the week of August 10," but the official Qwe​n organization still has no Qwen3.8 repository. Until the weights drop, nobody can host Qwen3.8-27B, so "free" is moot. Here are the three routes to free once the weights land (and what each one really costs), plus the models you can run free locally today.

No free API yet — the weights are the gate

Alibaba announced the Qwen3.8 family on August 3, 2026: the 2.4-trillion-parameter Qwen3.8-Max (about 95 billion active parameters, a 1M-token context, priced at $2 per million input and $6 per million output tokens on Alibaba Cloud Model Studio) and the dense, single-machine Qwen3.8-27B. Both were promised as open weights on Hugging Face and ModelScope the same week.

That promise is now two days overdue. I checked Hugging Face directly on August 12: the official Qwe​n organization has no Qwen3.8 repository of any kind. The Qwen3.8-27B-GGUF, -FP8, -NVFP4A16 and -NInfer repos that show up in model search are placeholder cards — zero weight files, single-digit download counts, model cards that literally read "reserved ahead of the Qwe​n/Qwen3.8-27B release." Do not treat those as evidence the model exists; treat them as people reserving parking spots.

Two things you cannot assume. First, the license: none has been named for Qwen3.8-27B. Recent Qwe​n open weights have shipped under the Tongyi Qianwen license, whose 100-million-monthly-active-users clause triggers commercial-license conversations at scale — Apache 2.0 is not a safe default. Second, the specs: Alibaba has published no benchmark table, context window, or confirmed hardware requirement for the 27B. Everything being repeated about it today is projection.

We covered the pre-drop checklist — what is confirmed, what is rumor, what to check before you download — in Qwen3.8-27B Open Weights: What We Know Before the Drop. This article is the part that piece did not cover: how "free" will actually work once the model ships.

After the weights drop: three routes to free, and what each really costs

"Free" is never free. All three routes to a free Qwen3.8-27B move the cost somewhere; the difference is where it lands. The 27B is a dense model — every one of its ~27 billion parameters is active on every token — so the costs below are about real hardware, not marketing.

Route 1: Run it yourself — free in cash, priced in hardware

The only route where "free" becomes literally true after the hardware is bought. Unsloth co-founder Daniel Han expects the 27B to run in about 17GB of VRAM on release — a target, not a shipped spec. Using the measured quantization ladder of Qwen3.6-27B as a proxy: Q4_K_M lands around 16GB, Q6_K around 21GB, FP8 around 27GB, and full BF16 around 54–56GB. The realistic minimum today is a 24GB card — RTX 3090, RTX 4090, or A6000-class — at Q4.

Timing matters: official weights with vLLM or SGLang usually work within days of a drop, but community GGUF and AWQ quantizations trail by one to two weeks, so an Ollama-style "pull and run" will not exist on release day. The hidden costs are power, a quiet machine, and your time. The win is privacy — nothing leaves your box — and a marginal cost of zero that compounds if you use the GPU for other models too.

Route 2: Hosted free tiers — free in cash, priced in limits

Once the weights are out, expect an evaluation quota from Alibaba Cloud and free tiers from the usual serverless hosts. The pattern to expect is the one Alibaba already runs for Qwen3.8-Max: a 1M-token new-user quota, Singapore region only, valid 90 days, no rollover — and reasoning tokens bill as output, so a model that reasons by default can burn the whole allowance in a weekend of prototyping. What you actually pay is rate limits, a regional lock, an expiry date, and your prompts going to a third party. A free tier is an on-ramp to evaluate the model, not a place to deploy it.

Route 3: OrcaRouter's free tier — planned, not live

Because the search query is literally "free," we are going to be explicit: OrcaRouter plans a free tier for Qwen3.8-27B once the weights ship. It is not live. There is no endpoint to call, and I am not going to paste a placeholder example. We will announce it when there is something to announce. What we do host today is Qwen3.8-Max at $2/$6 per 1M tokens with no markup — and the 27B is not on the platform, because nobody can serve it yet.

What you can use free right now

If your need is "a 27B-class open model I can run without paying," you do not have to wait. The honest same-level answer is Qwen3.6-27B, the direct predecessor to the model you are waiting for.

Qwen3.6-27B is Apache 2.0, a dense 27.8B model with 262K context and native text, image, and video input. A Q4_K_M GGUF is about 16.5GB and runs at roughly 25 tokens per second on a 24GB consumer GPU (16–18 tokens per second on an M4 Max). Vendor-reported numbers from its model card: SWE-Bench Verified 77.2, MMLU-Pro 86.2, AIME 2026 94.1, GPQA Diamond 87.8, and MMMU 82.9. If Qwen3.8-27B is "a 27B," this is the 27B you can actually touch today — same family, same dense design, already open.

Nemotron 3.5 Lightning 30B A3B is a different shape, also free: released August 11, 2026 under NVIDIA's OpenMDW 1.1 license. It is a 30B-total, ~3B-active Mixture-of-Experts model with a hybrid Mamba-2 architecture and up to 1M context — NVFP4 checkpoints are about 21.6GB and a Q4 GGUF about 25GB. It needs a 24GB+ card, because all 30 billion parameters sit in memory even though only about 3 billion activate per token. First reports put it near 45 tokens per second on a single GPU. It is built for the agent execution layer — tool calls, validation, formatting — not frontier reasoning. If your load is high-volume agent grunt work, it can beat a dense 27B at your actual job.

And if what you want specifically is a free hosted API today rather than a local install, the only honest "free" is Alibaba's Qwen3.8-Max quota: 1M tokens, Singapore region, 90 days. It is not the 27B, and it is an evaluation allowance, not a service — use it to try Qwen3.8 behavior now, then move.

When this advice is wrong

If you already own a 24GB+ GPU, the "wait for free tiers" framing is wrong for you. The day the weights drop, vLLM on the official weights is your free tier; you would be waiting for convenience you do not need. Hold out about two weeks only if you specifically want the polished GGUF quantization ladder.

If you are prototyping against an API contract, the hosted free quotas (once the 27B has them) may cost less than your time — even with the 90-day expiry and the region lock, renting a GPU box for a week of prototyping is usually the more expensive route.

If the license turns out to be Tongyi Qianwen rather than Apache, "free weights" will not mean free to commercialize above the 100-million-MAU threshold. Read the LICENSE file in the actual repository before you build anything on it — that is the single most common way "free open-weights" becomes an invoice.

And if Alibaba slips the date again — it is already two days past the promised window — every week that passes makes Qwen3.6-27B the correct call rather than the stopgap.

Bottom line

There is no free Qwen3.8-27B API because there is no Qwen3.8-27B anywhere. When the weights ship, the three free routes are self-hosting (pays in hardware), hosted free tiers (pays in limits), and OrcaRouter's planned free tier — which is not live, and we will say so when it is. Today, the closest free thing is Qwen3.6-27B on your own hardware. Waiting costs you nothing except the 27B; running the predecessor costs you nothing except a GPU you can put to work on everything else in the meantime.