A week after Qwen3.8-27B dropped (Aug 14, Apache 2.0, vision-capable, 27B dense, 262K native context), the community has gone from "is it good?" to "how do I actually run it without it spiraling?" The launch-week debate was quants and KL divergence. The week-two harvest is different: real hands-on reports on the model's default behavior — and the most interesting one is a genuinely funny bug that's actually a config problem.
1. The default is xhigh reasoning effort — and it's hilarious (until it costs you 21 minutes)
Simon Willison, who ran the model on both a 128GB M5 Max MacBook Pro and a DGX Spark (LM Studio, plus llama-server direct), opened his Aug 16 post with the line everyone's quoting: "Qwen3.8 comes with official support for reasoning_effort... xhigh (default). This is a hilarious default. It's absolutely not a good way to run the model."
The evidence is a masterpiece of overkill. His first prompt — an SVG of a pelican riding a bicycle — came back after 21 minutes, using 22,276 reasoning tokens to produce 3,223 output tokens. The result was genuinely the best local-model SVG he'd generated (correct bicycle frame, legs on both sides of the bike, wings touching the handlebars). "Was that worth waiting 21 minutes for? Absolutely not."
Then the same prompt with reasoning off: 3,715 tokens, 137 seconds — under two and a half minutes.
And the one that made the point stick: "draw an svg of a circle" at the default xhigh setting produced a reasoning trace that read like a design brief ("geometric 'circle study'... pulsing glow... Bauhaus/compass-drawing vibe...") and delivered — "several minutes later" — a beautiful animated circle that was "entirely not what I had asked for."
Two practical takeaways buried in there:
- He hit LM Studio's 8,192-token default context limit immediately — the model used all of it thinking about mundane problems. Loading with the full 262,144 max context fixed it. If you're testing the 27B on a laptop, set the context to the model's native limit or you'll see truncated/spiraling outputs and blame the model for a config choice.
- The reasoning_effort parameter is the single most important setting on this model. xhigh is a default for the Qwen Cloud API's use case, not for local consumer hardware. Set it to medium or low (or off) for interactive work.
2. The speed wave: 2x on a 5090, Cerebras on the way
Where launch week was about quants, week two is about throughput — the community isn't waiting for Qwen to tune defaults, it's shipping fixes:
- balto-speedrunner (GitHub, Aug 16): a repo claiming Qwen 3.8 27B at 2x speed on an RTX 5090 — a snapshot in the "make the 27B run fast on consumer Blackwell" race. Expect a flood of speculative-decoding/attention-optimization forks targeting the same hardware class.
- Cerebras (Aug 16): official "Qwen3.8 27B Coming to Cerebras" — inference at 40GB/s-class memory bandwidth with no VRAM constraint. For anyone who wants the full 262K context without quantization trade-offs, Cerebras (and equivalents like Groq's roster) is the zero-VRAM path.
- Show HN (Aug 17): a hosted Qwen3.8-27B API at 140 tok/s on one GPU — someone is already selling the "one GPU" story as a product, which is the fastest possible signal that this model's deployment lane (27B dense on a single accelerator) is the one the market actually wants.
The pattern is visible: the 27B is becoming the default "runs on my hardware" model, and the week-two activity is all about the two things that decide whether that default is pleasant: reasoning-effort defaults and tokens-per-second.
3. The "endless whiteboard" demo: what the model's overthinking is actually good for
Kenton Varda (Cloudflare) shipped a fun one: "Qwen 3.8 27B in Cloudflare OS made me a collaborative whiteboard" — using the model as the local "brain" of a shared canvas. It's a small demo, but it's the right shape: a 27B vision model running locally as an always-on ambient agent. That's the use case the xhigh default accidentally serves — when you want the model to keep generating ideas, the overthinking ceases to be a cost and becomes the product.
4. What to watch in week three
- The "I changed the default" crop: now that reasoning_effort is common knowledge, expect a wave of retests with medium/low showing 5-20x wall-clock improvements on simple tasks. The honest benchmarks of the model are being written against the default, so the next round of "Qwen 27B is fast" posts will quote the tuned config.
- NVFP4 / Balto / driver-level speedups landing on 5090-class hardware — watch for verified (not vendor-claimed) throughput numbers.
- Long-context agentic tests with the 262K context and reasoning set to low: the launch-week thread flagged quantized models losing focus after long context; week three should tell us whether official weights + low reasoning makes the 27B a genuine long-horizon agent core.
- Cerebras/Groq availability turning the "local model" story into "frontier-speed open model without a GPU."
The bottom line
Week-two verdict on Qwen 3.8-27B: the model is excellent and the default is wrong. The overthinking isn't a defect — it's a config choice shipped for cloud API economics, and Simon Willison's 21-minute pelican is the funniest documentation of that since the model launched. The community's response is already the right one: tune the reasoning effort, widen the context window, and race to make it fast on consumer hardware. If you're trying the 27B this week, skip the benchmark arguments — spend the first five minutes changing the default and setting the context to 262K. That's the single highest-leverage move the launch threads produced.
AI Tools is a practical briefing on the AI tools, models, and licenses shaping developer workflows.
Top comments (0)