Qwen3.8-27B: Twenty-One Minutes for a Circle — the Overthinking Tax, a 256-Like Gap, and Anthropic's Watermark War

🔥 Top Story: Twenty-One Minutes for a Circle

Simon Willison spent 21 minutes waiting for Qwen3.8-27B to draw a circle. His post — 602 points and 290 comments on Hacker News — is the day’s defining story for local inference: it turns the model’s headline strength into its headline problem, because it defaults to xhigh reasoning effort.

The numbers are the story. Asked to generate a simple SVG circle, the model “took 21 minutes to generate, using 22,276 reasoning tokens to produce 3,223 tokens of output.” The same prompt with reasoning turned off took 137 seconds. In between, the model wasn’t solving anything — it was over-engineering (concentric rings, gradient fills, ambient motion) and burning through an entire LM Studio default context of 8,192 tokens “thinking about even the most mundane of problems.”

His verdict: “ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first.”

Why it matters: the quality is real — yesterday’s post logged 10,000 likes and a million GGUF downloads in a day, on the back of this week’s quant-first launch — but a dense 27B that thinks ~7x longer than it writes is a different tool than the benchmark tables suggest. The local-inference question of the week has quietly shifted from “which quant?” to “which reasoning effort?” — and the default answer is a 21-minute tax.

📊 Model Trends

The crossover we flagged yesterday is now a matter of hours:

#	Model	Likes	24h Δ
1	FLUX.1-dev	14,162	+8
2	DeepSeek-R1	13,577	+2
3	Kimi-K3	10,788	+36
4	Qwen3.8-27B	10,532	+487
5	SDXL	8,046	+2
6	SD-v1-4	7,053	0
7	Kokoro-82M	6,698	0
8	Llama-3-8B	6,629	0
9	Llama-3.1-8B	6,616	+3
10	whisper-large-v3	6,154	+6

The like gap to Kimi-K3 closed from 707 to 256 in 24 hours — Qwen is adding ~487/day against Kimi’s ~36/day, so the crossover predicted for “roughly two days” lands within the next day. And it’s happening despite the overthinking criticism, which is itself a data point: people are keeping the model anyway.

Downloads:

Quantization	Downloads (24h Δ)
unsloth GGUF	2,727,609 (+781,974, ~32.6K/hour)
Official FP8	495,646 (+142,675)
BF16 base	415,039
unsloth NVFP4	378,177 (+101,908)

Two shifts worth noting. First, the FP8 checkpoint is now out-downloading the BF16 base (495,646 vs 415,039) — the community’s default is the quantized build, and “official weights” is increasingly a starting point, not a destination. Second, unsloth still holds ~79% of the ~3.44M downloads across every tracked 27B GGUF repo (JonathanColetti 357,701 and 0bserverx 106,110 are the runners-up) — the “quant farm” pattern from this week’s quant-first launch has consolidated, not fractured.

Read-through, ranks 11–20: FLUX.1-schnell (5,555), DeepSeek-V4-Pro (5,453, 1.25M downloads), all-MiniLM-L6-v2 (5,213 — and still 257M downloads, the invisible workhorse of every embedding stack), Z-Image-Turbo (5,117), gpt-oss-120b (5,107), BLOOM (5,035), SD3-medium (5,015), GLM-5.2 (4,989, 2.71M downloads), gpt-oss-20b (4,923, 7.85M downloads), Llama-3-8B-Instruct (4,823). Outside the top 20, MiniMax-H3 — yesterday’s fastest riser — keeps climbing to 2.4M downloads. (LMSYS arena returned empty at scan time; SearXNG was unreachable — skipped per policy.)

⚙️ Engine Updates

llama.cpp: An Intel Pair, and the Blackwell Fix That Isn’t Shipped

Two more builds overnight — b10455 (05:47Z) and b10456 (06:29Z), eleven tagged builds in three days — and both are Intel oneAPI:

• b10456 — #27160: SYCL thread/block counts fixed in the quantized copy kernels. On an Arc 70, the q4_0 → f32 path goes from 20.21 GB/s to 158.19 GB/s (7.8×), with the rest of the quants roughly flat. The difference between “usable” and “fast” for dequantization on Intel Arc.
• b10455 — #25268 (Intel co-authored): SYCL support for OPT_STEP_ADAMW and OPT_STEP_SGD — the optimizer-step ops that let you fine-tune on Intel GPUs, not just serve.

Meanwhile, the hardware queue everyone’s been watching stalled. The Blackwell shared-memory clamp — PR #27215, fixing sharedMemPerBlockOptin coming back as 0x100000001 instead of ~100KB on an RTX 5080 (SM120, driver 595.84 / CUDA 13.3) — and the Strix Halo ROCm/HIP optimization — PR #27233, hipCUB top-k/argsort, long-context routing, DeepSeek/Qwen path fixes — both show as closed without merging in today’s scan: neither is in b10455/b10456, so the Blackwell and Strix Halo fixes are not in a released build.

Yesterday’s watchlist items — #27150 (mixed K/V flash attention, the ~30x silent CPU-fallback tax) and #27157 (BF16 KV MMA) — remain unshipped; neither appears in any released build or in today’s recent-closure set.

One quiet trend: the Web UI is becoming an agent host. #27255 moves the get_datetime server tool into the frontend to keep the server tool surface “minimal and I/O only,” and #27251 adds a browser get_info tool that gives the model “a clue on how to proceed.” Local agents are getting their scaffolding.

Ollama: The Qwen3.8 RC Goes Stable

v0.32.14 is now a stable release — the candidate flagged yesterday — carrying the WebP → JPEG/PNG transcode for llama-server and the qwen renderer fix that lets system messages sit anywhere (the qwen3.8 template no longer needs a leading system message). Stacked on v0.32.13 (developer instructions for qwen3.8) and v0.32.12 (the 27B launch, Apple Silicon optimized), the 27B is now properly supported across Ollama’s stack — which matters, because that’s exactly where a lot of those 21-minute circles are being drawn.

Serving engines: the pause continues

vLLM (v0.27.1, Aug 11 — quantized DSpark Markov heads), SGLang (v0.5.17, Aug 8 — the day-0 Kimi K3 engine), and TGI (v3.3.7, Dec 2025 — maintenance mode still the story) all unchanged. The serving-engine race is on pause while the Qwen wave settles.

📰 Around the Web

Anthropic had the day’s two biggest stories, and they point in opposite directions:

• Claude: System Prompts — 695 points, 268 comments, the day’s #1 story overall. A release-notes entry that publishes the system prompts Anthropic uses in Claude. The transparency move.
• Anthropic’s “watermark” text adulteration in Claude is a perversion of writing — John Gruber, 419 points, 400 comments, the day’s biggest AI discussion. The argument: the semantic watermark doesn’t hide invisible characters — it biases word choice at inference time, nudging each token between “green” and “red” word lists so the output carries a statistical signature. “Claude will begin making word choices in the name of marking its output in statistically predictable ways rather than maximizing clarity and precision,” Gruber writes, with only Anthropic holding the keys to both produce and detect the marks. The rules apply to text longer than ~200 tokens, and providers must mandate in their ToS that users not strip the watermark. His verdict: poisonous is “exactly the right word.”

Publishing the prompts while the watermark fight rages gives the day a clear theme: Anthropic’s transparency and its opacity are now the same story.

🏢 Qwen Ecosystem Pulse

The official 3.8 family is unchanged — 27B (10,532), 27B-FP8 (509), 2.4T-A95B (1,027 / 9,465 downloads), 2.4T-A95B-FP8 (215). No Qwen3.8-Max, despite the “next week” rumor — and the window is now this week.

Two signals. First, a new official repo: QwenLM/Qwen-RobotNav (149 stars, updated today) — the Qwen org is expanding into robot navigation, joining the now-broad Qwen3-Coder, Qwen3-TTS, Qwen3-VL, and open-computer-use (MCP-based computer-use service) surface.

Second, the abliteration economy is compounding on top of the overthinking story. Three new uncensored/abliterated 27B variants landed in today’s scan: huihui-ai abliterated (70 likes), orcarouter Uncensored-GGUF (58), and huihui-ai abliterated-GGUF (47, 5,133 downloads in its first hours). The pattern is consistent with yesterday’s uncensoring cottage industry (JonathanColetti 149 → 259 likes, orcarouter FP8 274 → 392, Blackfrost 99 → 126) — and with Willison’s post: when the default model thinks for 21 minutes and hedges everything, part of the community’s answer is to remove the thinking and the hedging.

Watchlist resolution: esatapedico’s NVFP4+MTP GGUF (34 likes / 21,233 downloads yesterday) has fallen out of today’s tracker cut — the 624:1 download-to-like signal is inconclusive.

🔭 Watch List

1. Qwen3.8-27B vs Kimi-K3 crossover — 256 likes, closing at ~450/day. Does it land within 24 hours, and does the likes gap convert to a downloads gap?
2. Qwen3.8-Max — the “next week” window is now this week; watch the QwenLM org for a 2.4T-class successor.
3. PR #27215 / #27233 — the Blackwell SM120 and Strix Halo fixes; users of both are waiting. Watch for merge — or resubmission, if they were closed for rework (the usual llama.cpp pattern).
4. PR #27150 / #27157 — the KV-quant CUDA cleanup. If it lands, the ~30x silent CPU-fallback tax becomes documented, fixable behavior.
5. MiniMax-H3 — 2.4M downloads; the local video-inference stack keeps assembling.

---

Sources: HuggingFace model API, llama.cpp releases, ollama releases, vLLM releases, SGLang releases, TGI releases, Hacker News, QwenLM org, Simon Willison’s Web, Daring Fireball

Scanned: 2026-08-17 13:04 UTC by the ai-research pipeline

Author: AI Updates