Checked August 14, 2026. Qwen3.8-27B is now a real downloadable Qwen checkpoint, not just a rumored smaller sibling of Qwen3.8-Max. The new Qwen/Qwen3.8-27B repository identifies Qwen as the author, reports 27,781.4 million parameters, uses the qwen3_5 architecture tag, is exposed as image-text-to-text, and carries an Apache-2.0 license. An official FP8 build and community MLX/GGUF conversions appeared on the same day. [S1–S6]
That launch speed makes the model immediately relevant to Apple Silicon. The current official model card now includes a Qwen3.8-27B benchmark table and specifies 262,144 native context tokens, with an extension to 1,000,000 tokens described through RoPE/YaRN configuration. Those values belong to the downloadable checkpoint; the hosted Qwen3.8 service is a different deployment surface. Independent Qwen3.8-27B Mac benchmarks and reproducible runtime comparisons are still missing.
Launch-day facts
The key launch questions at a glance, with the source noted per row:
| Item | Verified status |
|---|---|
| Official Qwen checkpoint | Yes: Qwen/Qwen3.8-27B. [S1] |
| Parameter metadata | 27,781.4M. [S1] |
| License | Apache 2.0. [S1, S2] |
| Multimodal task tag | image-text-to-text. [S1] |
| Official FP8 | Yes. [S2] |
| MLX | 4-bit, 8-bit and BF16 conversions already exist. [S3–S5] |
| GGUF | Unsloth package available. [S6] |
| Official model-card benchmark table | Available; independent reproduction still open. |
| Native context | 262,144; extendable to 1,000,000 with documented configuration. |
| oMLX support | Strong compatibility signals, but needs a direct Qwen3.8-specific test. |
Do not confuse the 27B checkpoint with Qwen3.8-Max-Preview
Alibaba launched qwen3.8-max-preview through its Token Plan on July 19, 2026. QwenCloud lists that hosted model with a one-million-token context window. [S7–S9]
Qwen3.8-27B is a different product surface: downloadable weights, a 27B parameter class and Apache 2.0. The Max Preview service’s context window, pricing, infrastructure and vendor positioning should not be pasted onto the local 27B checkpoint.
This distinction matters for Mac users. One is a hosted service; the other can be quantized and served on your own machine. [S1–S6]
How much unified memory does a 27.78B model need?
A useful first-pass calculation is the theoretical raw weight storage:
storage ≈ parameter count × bits per weight ÷ 8
Using the repository’s 27.7814B parameter metadata:
| Precision | Theoretical raw weights |
|---|---|
| BF16 / 16-bit | ~55.6 GB |
| FP8 / 8-bit | ~27.8 GB |
| 8-bit quant | ~27.8 GB |
| 4-bit quant | ~13.9 GB |
These are calculated estimates, not measured runtime memory. Real usage is higher because a running model also needs quantization metadata, runtime structures, KV cache, activations, allocator headroom and memory for macOS and the surrounding agent stack.
Practical Mac tiers
16 GB: A 4-bit model is already too close to the raw-weight floor for comfortable agent use.
24 GB: 4-bit is plausible, but long contexts and parallel tasks can consume the remaining headroom quickly.
32 GB: The natural starting point for 4-bit. Eight-bit weights alone are too close to total memory for comfort.
48 GB: Plenty of space for 4-bit and a more plausible target for 8-bit.
64 GB: Eight-bit becomes much easier to accommodate. BF16 still puts roughly 55.6 GB into weights before runtime overhead.
96 GB+: BF16 becomes technically reasonable, although it is unlikely to be the best value for most local-agent workloads.
Which MLX build should you start with?
The fastest practical answer on launch day is 4-bit MLX. mlx-community/Qwen3.8-27B-4bit already exists, and its theoretical weight floor is roughly 13.9 GB. [S3]
That leaves much more unified memory for a real agent workload than an 8-bit build. Eight-bit becomes interesting when you have the memory and want to measure whether the extra precision improves your actual tasks. Without a Qwen3.8-specific quantization study, claiming a fixed quality gain would be speculation.
What about oMLX?
oMLX already has extensive support history for Qwen3.5/Qwen3.6-family models, including continuous batching, SSD caching and Qwen-specific fixes. [S17–S25] Qwen3.8-27B retains the qwen3_5 architecture tag and has immediate MLX conversions, so there is a credible compatibility path.
But launch-day compatibility should still be treated as a testable hypothesis rather than a finished fact. The predecessor family exposed runtime-specific edge cases around vision fallback, thinking controls, MTP, quantization and long-context throughput. [S21–S25]
A useful smoke test for Qwen3.8-27B in oMLX should separately validate:
- basic OpenAI-compatible text generation,
- repeated tool calls,
- image input if you need vision,
- memory growth at longer contexts,
- MTP or DFlash only after Qwen3.8-specific support is confirmed.
The last point is important. oMLX’s DFlash documentation maps target models to specific draft checkpoints. The Qwen3.6 draft model should not be assumed to work with Qwen3.8. [S19]
Is Qwen3.8-27B actually better than Qwen3.6-27B?
The official model-card table provides an initial vendor-reported comparison, but it is not enough to answer the local-agent question independently.
The new checkpoint is official, its model-card evaluation is now accessible, and the local ecosystem is real. [S1–S6] The remaining quality question is whether independent third-party and local Mac tests reproduce those results.
The official Qwen3.8 table is a useful starting point, but it is vendor-reported rather than an independent Mac evaluation. The predecessor is also a warning against simplistic score reading: Qwen3.6 community reproduction work showed that repository-edit tools, test conditioning and the overall agent harness can dramatically alter SWE-bench outcomes. [S26–S28]
For a meaningful generation-to-generation test, hold constant:
- quantization class,
- system prompt,
- tool schema,
- context budget,
- sampling settings,
- agent harness,
- task set,
- number of retries.
The benchmark that matters for a local Mac agent
The official table is not enough for a local-Mac verdict. A useful launch review should publish four additional groups of numbers:
Decode throughput: tokens per second at 4-bit and 8-bit.
Prefill throughput: especially at 8K, 32K and 64K contexts.
Peak unified memory: measured during a full agent run rather than model load alone.
Agent reliability: file editing, shell use, structured tool calls, research and multi-step execution.
oMLX’s own Qwen3.6 history shows why runtime version must be recorded: long-context prefill performance has changed materially between releases. [S22]
Practical status
Qwen3.8-27B has been available as an official Qwen checkpoint on Hugging Face since August 14. For local Mac users, the useful fact is its size class: 4-bit MLX, 8-bit and BF16 conversions, and a GGUF package are already available, while independent Mac and agent tests are still missing. [S1–S6]
The open question is independent local quality, not availability. Until reproducible Mac and runtime results land, the best starting configuration is 4-bit MLX on a Mac with roughly 24–32 GB of unified memory or more, based on the weight-size math and the need for runtime headroom—not on a measured quality claim.
Source note
References [S1]–[S35] point to the list below. Time-sensitive facts were checked on August 14, 2026.
Sources
[S1] https://huggingface.co/Qwen/Qwen3.8-27B [S2] https://huggingface.co/Qwen/Qwen3.8-27B-FP8 [S3] https://huggingface.co/mlx-community/Qwen3.8-27B-4bit [S4] https://huggingface.co/mlx-community/Qwen3.8-27B-8bit [S5] https://huggingface.co/mlx-community/Qwen3.8-27B-bf16 [S6] https://huggingface.co/unsloth/Qwen3.8-27B-GGUF [S7] https://modelstudio.alibabacloud.com/intl/blog/model-studio-token-plan-individual/ [S8] https://www.alibabacloud.com/help/en/model-studio/token-plan-harness-tool [S9] https://docs.qwencloud.com/developer-guides/getting-started/text-generation-models [S10] https://qwen.readthedocs.io/en/stable/inference/transformers.html [S11] https://huggingface.co/Qwen/Qwen3.6-27B [S12] https://huggingface.co/Qwen/Qwen3.6-27B-FP8 [S13] https://huggingface.co/Qwen/Qwen3.5-27B [S14] https://huggingface.co/Qwen/Qwen3.6-35B-A3B [S15] https://arxiv.org/abs/2505.09388 [S16] https://github.com/QwenLM/Qwen3.6 [S17] https://github.com/jundot/omlx [S18] https://github.com/jundot/omlx/releases [S19] https://github.com/jundot/omlx/blob/main/docs/experimental/dflash_mlx_integration.md [S20] https://github.com/jundot/omlx/discussions/945 [S21] https://github.com/jundot/omlx/issues/1097 [S22] https://github.com/jundot/omlx/issues/2155 [S23] https://github.com/jundot/omlx/issues/1378 [S24] https://github.com/jundot/omlx/issues/1464 [S25] https://github.com/jundot/omlx/issues/1556 [S26] https://github.com/QwenLM/Qwen3.8/issues/179 [S27] https://github.com/QwenLM/Qwen3.8/issues/178 [S28] https://github.com/QwenLM/Qwen3/discussions/1846 [S29] https://github.com/QwenLM/qwen-code/issues/3878 [S30] https://www.airealist.ai/p/qwen-38-soon-is-not-a-date [S31] https://gigazine.net/gsc_news/en/20260721-qwen3-8 [S32] https://www.nxcode.io/resources/news/qwen3-8-benchmarks-coding-agent-evaluation-guide-2026 [S33] https://benchlm.ai/compare/qwen3-5-27b-vs-qwen3-8-max-preview [S34] https://modelfit.io/ [S35] https://aiwiki.ai/wiki/qwen3_6
Frequently Asked Questions
Is Qwen3.8-27B an official Qwen checkpoint?
Yes. Qwen publishes it as Qwen/Qwen3.8-27B on Hugging Face. The model card lists a 27B parameter class, Apache 2.0, image and video understanding, and 262,144 native context tokens. [S1]
How much unified memory does Qwen3.8-27B need on a Mac?
The theoretical raw weights are about 13.9 GB at 4-bit, 27.8 GB at 8-bit and 55.6 GB at BF16. Runtime structures, KV cache, vision processing and macOS add overhead, so 24–32 GB is the more realistic 4-bit starting tier. [S1][S3–S5]
What is the context window?
The official model card specifies 262,144 native tokens. Qwen describes extending this to 1,000,000 tokens through RoPE/YaRN configuration; that is not the same as an automatically available 1M window in every local runtime. [S1]
Is Qwen3.8-27B better than Qwen3.6-27B?
The model card now includes an official comparison table, but independent Mac and runtime tests are still missing. A defensible comparison must hold quantization, harness, tool schemas, context and tasks constant. [S1][S26–S28]
Does Qwen3.8-27B already work with oMLX?
The qwen3_5 architecture and immediate MLX conversions are strong compatibility signals. A Qwen3.8-specific oMLX release and a direct tool/vision smoke test still need to be confirmed separately. [S1][S3–S5][S17–S25]