Chat model
Qwen3.8 27B on Apple silicon Macs
Released
Strong chat, writing and coding. Smarty Pants serves a pinned, quantized 3.8-bit MLX build on qualified Macs with 32 GB of memory or more, through an OpenAI-compatible API.
$0.20
$0.20 per million input tokens, $1.20 per million output tokens
Listed price on November 19, 2026. Current price: GET /v1/models.
Qwen3.8 27B at a glance
What Smarty Pants serves, from its public catalog, and what the model's developer publishes about it.
- API name
smartypants/qwen3.8-27b- Endpoint
- POST /v1/chat/completions
- Build served
- 3.8-bit MLX (qwen3.8-27b-mlx-dwq3.8-r9fa7ab0)
- Mac memory needed
- 32 GB or more
- Context on Smarty Pants
- 16,384 tokens
- Max output
- 4,096 tokens
- Developer
- Qwen (Alibaba Cloud)
- Parameters
- 27B
- License
- Apache-2.0
- Upstream context
- 262,144 tokens (extensible up to 1,000,000)
- First released
- August 2026
- Status
- Released
Smarty Pants catalog checked November 19, 2026. Upstream facts from the developer's model card (sources below).
What Qwen3.8 27B is good for
Dense vision-language model aimed at coding, professional work, research and long-horizon agentic tasks, with native image and video understanding (including hour-scale videos).
Thinking is on by default but can be disabled per request, reasoning depth can be tuned with a reasoning_effort setting, and preserve_thinking keeps reasoning from earlier turns in context.
The upstream model also accepts images. Smarty Pants serves it for text in, text out.
Use it for
- Writing and editing that needs more care
- Coding help
- Agent steps with tool calls
Look elsewhere for
- Inputs longer than 16,384 tokens
- Latency-critical chat (it is slower than the small models)
Measured on a real Mac
A standard 4-bit build ran at 35.7 tokens a second with 18.26 GB peak memory (4,096-token prompt) on an Apple M5 Pro Mac mini. Smarty Pants now serves a smaller 3.8-bit build.
Measured 2026-11-02 by the Smarty Pants model factory. Full method and results: Smarty Pants Mac benchmarks on Hugging Face. Live speed depends on the Mac serving your request.
How to call Qwen3.8 27B
Smarty Pants works with OpenAI's libraries: change the base URL and the key, and keep the rest. Get a key at smartypants.network/api.
from openai import OpenAI
client = OpenAI(base_url="https://smartypants.network/v1", api_key="YOUR_SMARTY_PANTS_KEY")
reply = client.chat.completions.create(
model="smartypants/qwen3.8-27b",
messages=[{"role": "user", "content": "Hello!"}],
)
print(reply.choices[0].message.content) Before sending real traffic, check is_ready for this model on GET /v1/models. A 429 or 503 means no Mac can take the request right now: wait the Retry-After seconds or use your fallback provider.
Which Macs can serve Qwen3.8 27B
Smarty Pants serves this model only on Macs with 32 GB of unified memory or more. These Apple silicon chips were sold with at least one memory size that big (memory options from Apple's tech specs).
| Chip | Memory sizes that fit | Memory sizes too small |
|---|---|---|
| M1 Pro | 32 GB | 16 GB |
| M1 Max | 32 GB, 64 GB | None |
| M1 Ultra | 64 GB, 128 GB | None |
| M2 Pro | 32 GB | 16 GB |
| M2 Max | 32 GB, 64 GB, 96 GB | None |
| M2 Ultra | 64 GB, 128 GB, 192 GB | None |
| M3 Pro | 36 GB | 18 GB |
| M3 Max | 36 GB, 48 GB, 64 GB, 96 GB, 128 GB | None |
| M3 Ultra | 96 GB, 256 GB, 512 GB | None |
| M4 | 32 GB | 16 GB, 24 GB |
| M4 Pro | 48 GB, 64 GB | 24 GB |
| M4 Max | 36 GB, 48 GB, 64 GB, 128 GB | None |
| M5 | 32 GB | 16 GB, 24 GB |
| M5 Pro | 48 GB, 64 GB | 24 GB |
| M5 Max | 36 GB, 48 GB, 64 GB, 128 GB | None |
| M5 Ultra | 96 GB, 256 GB, 512 GB | None |
| M6 | 32 GB | 16 GB, 24 GB |
No memory option is big enough on: M1, M2, M3.
Frequently asked questions
What is the API name for Qwen3.8 27B on Smarty Pants?
Use "smartypants/qwen3.8-27b" as the model with the OpenAI-compatible base URL https://smartypants.network/v1.
How much does Qwen3.8 27B cost on Smarty Pants?
The listed price on November 19, 2026 was $0.20 per million input tokens, $1.20 per million output tokens. Prices can change; GET https://smartypants.network/v1/models returns the current price, and you are not charged for a request Smarty Pants can't serve.
How much memory does a Mac need to serve Qwen3.8 27B?
Smarty Pants serves this 3.8-bit MLX build only on qualified Apple silicon Macs with 32 GB of unified memory or more.
What context length does Qwen3.8 27B have on Smarty Pants?
Smarty Pants accepts up to 16,384 tokens per request and writes up to 4,096. The upstream model card lists 262,144 tokens (extensible up to 1,000,000); Smarty Pants serves a smaller window so the model fits on Macs.
Is Qwen3.8 27B available right now?
On November 19, 2026 it was released. How many Macs have it online changes through the day; check is_ready on https://smartypants.network/v1/models before sending real traffic, and keep a fallback provider.
Can I send private data to Qwen3.8 27B on Smarty Pants?
Requests on the general network run on Macs other people own, and a serving Mac decrypts a request in memory to answer it. Keep secrets, personal and confidential data off the general network; companies can run work on their own Macs with a Group.
Sources
Other chat models
Try a pinned open model on real Macs.
See every model, its listed price and the code to call it. Smarty Pants works with the OpenAI SDK you already use: change the base URL and the key.
Requests run on Macs owned by other people. Owners agree not to read requests and the app doesn't show them, but a serving Mac decrypts each request in memory to answer it, so that is a rule and a design, not a technical guarantee. Don't send sensitive data on the general network.