Qwen3.8-27B Pricing: Local AI Costs & API Status

Alibaba’s Qwen team released Qwen3.8-27B on August 14 as an open-weight, 27-billion-parameter dense model for coding, professional work, research, and multimodal agents. The official BF16 and FP8 repositories are public and licensed under Apache 2.0.

The headline circulating on Hacker News calls it the “best local dense model yet.” That is a promising description, not an independently established fact. Qwen’s own results show large gains over Qwen3.6-27B, but buyers should reproduce them on private tasks before making a hardware or migration decision.

What Qwen3.8-27B adds

Qwen3.8-27B is a native vision-language model that accepts text, images, and video. It has a 262,144-token native context window, with documented extension to one million tokens, and supports low, medium, and xhigh reasoning effort. Thinking is enabled by default and can be disabled for faster direct responses.

Qwen publishes serving recipes for Transformers, vLLM, SGLang, and TokenSpeed. The official BF16 safetensors total about 55.56 GB, while the official FP8 safetensors total about 30.87 GB. Those are repository weight footprints, not complete VRAM requirements: runtime buffers, the vision projector, KV cache, batch size, and long context all add memory.

Community GGUF conversions appeared on launch day. A Q4_K_M file is about 15.9 GiB, plus a separate vision projector, making a 24GB-class machine a plausible test target. Actual speed, usable context, and quality loss depend on the runtime and quantization; launch-day community files are not production benchmarks.

Benchmark gains—and the caveat

In Qwen’s published table, Qwen3.8-27B scores 61.7 on SWE-bench Pro, ahead of Qwen3.6-27B at 53.5 and the listed Opus 4.6 Max result at 53.4. On Terminal Bench 2.1, it scores 73.0 versus 63.4 for Qwen3.6-27B and 78.2 for Opus 4.6 Max.

These are provider-reported results. Qwen says most SWE-bench Pro models were run through the Claude Code harness on a corrected task set, while the Opus result came from its official report. That is useful evidence for an evaluation shortlist, but it is not a neutral, independently replicated leaderboard.

Pricing impact: open weights are not free inference

The exact Qwen3.8-27B hosted price is not published. Qwen’s model card says its Cloud endpoint—with default 1M context and built-in tools—is “coming soon.” The live table and chart above therefore use currently tracked Qwen and frontier API baselines; none is a proxy price for the new 27B model.

Self-hosting replaces per-token billing with hardware, electricity or GPU rental, engineering time, monitoring, and idle capacity. It is strongest when traffic is steady, privacy or offline operation matters, and the model meets an acceptance test without expensive retries. A hosted API usually wins for bursty workloads or teams that cannot keep GPUs well utilized.

Who benefits—and who should wait

Local AI users gain the most. A dense 27B model is easier to reason about and deploy than a huge mixture-of-experts release, while Apache 2.0 permits broad commercial use. Coding teams, private document workflows, and multimodal agents now have a serious local candidate with adjustable reasoning depth.

Teams needing audited uptime, predictable throughput, or a published unit price should wait for managed hosts to add the exact snapshot. Buyers should also avoid treating a one-million-token context claim as a practical desktop target until KV-cache memory, latency, and accuracy are measured at that length.

What to do now

1. Start with FP8 on supported data-center GPUs or a conservative GGUF quantization on local hardware.
2. Test coding, tool use, vision, latency, memory, and power on production-shaped tasks—not only public benchmarks.
3. Compare thinking off, low, and medium before accepting the default xhigh token and latency overhead.
4. Calculate full monthly infrastructure cost and divide it by accepted outputs, including retries and operator time.
5. Keep the exact model commit and quantization pinned; launch-day artifacts can change quickly.

Compare current routes on the Qwen pricing page, Novita endpoint guide, and OpenAI pricing page, then model your traffic in the token calculator. Our API versus self-hosting break-even guide provides the decision framework.

Teams that want a managed Qwen test before buying hardware can check Novita’s current Qwen catalog. Confirm that the exact Qwen3.8-27B snapshot is listed before comparing results.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect this analysis.

Sources: Qwen’s official Qwen3.8-27B model card, official FP8 release, Alibaba Cloud’s Model Studio catalog and pricing page, community GGUF files, and the Hacker News discussion.