256 Likes, and Gone — Qwen3.8 Passes Kimi-K3, llama.cpp Ships v0.1.2, and Windows ARM64 Gets CUDA
Qwen3.8-27B closes the 256-like gap and takes
🔥 Top Story: 256 Likes, and Gone
The crossover yesterday’s post called “a matter of hours” has landed: Qwen3.8-27B is now #3 on HuggingFace with 10,960 likes (+428 in 24h), past Kimi-K3 at 10,817 (+29). The 256-like gap closed overnight and flipped into a 143-like lead.
The four-day arc is now complete: quant-first launch (08-15) → 10K likes in a day (08-16) → 21-minute circle and the 256 gap (08-17) → crossover (08-18).
Why it matters: this happened despite the overthinking criticism. Simon Willison’s post — 21 minutes and 22,276 reasoning tokens to draw a circle — was the week’s loudest warning, yet likes kept climbing at ~430/day. The community’s answer wasn’t to abandon the model; it was to keep it and tune it. Yesterday’s default question had already shifted from “which quant?” to “which reasoning effort?” Today’s data is the answer: low effort, rerun if it thinks too long — and keep the model.
📊 Model Trends
The crossover landed overnight:
| # | Model | Likes | 24h Δ |
|---|---|---|---|
| 1 | FLUX.1-dev | 14,172 | +10 |
| 2 | DeepSeek-R1 | 13,580 | +3 |
| 3 | Qwen3.8-27B | 10,960 | +428 |
| 4 | Kimi-K3 | 10,817 | +29 |
| 5 | SDXL | 8,047 | +1 |
| 6 | SD-v1-4 | 7,054 | +1 |
| 7 | Kokoro-82M | 6,705 | +7 |
| 8 | Llama-3-8B | 6,629 | 0 |
| 9 | Llama-3.1-8B | 6,621 | +5 |
| 10 | whisper-large-v3 | 6,159 | +5 |
Qwen added 428 likes against Kimi’s 29 — a 15:1 rate on the day the gap closed. Everything else on the leaderboard moved in single digits; the crossover is the only structural change in the top 10.
Downloads:
| Quantization | Downloads (24h Δ) |
|---|---|
| unsloth GGUF | 3,561,466 (+833,857, ~34.7K/hour) |
| Official FP8 | 741,011 (+245,365) |
| BF16 base | 665,513 (+250,474) |
| unsloth NVFP4 | 523,919 (+145,742) |
FP8 still leads BF16, but the gap narrowed from ~80.6K to ~75.5K — BF16 is now downloading faster (+250,474 vs +245,365). And the downloads-side crossover happened quietly: unsloth’s GGUF repo alone (3,561,466) has passed Kimi-K3’s entire download count (2,226,898). unsloth holds ~76% of the 4.69M across the 10 tracked 27B GGUF repos; runners-up are JonathanColetti (558,767, +201,066), 0bserverx (150,262, from 106,110), and bartowski (105,006).
Read-through, ranks 11–20: FLUX.1-schnell (5,564), DeepSeek-V4-Pro (5,457), all-MiniLM-L6-v2 (5,216 — still 258M downloads, the unkillable embedding workhorse), Z-Image-Turbo (5,119), gpt-oss-120b (5,109), BLOOM (5,035), SD3-medium (5,015), GLM-5.2 (5,000, 2.73M downloads), gpt-oss-20b (4,926, 7.8M downloads), Llama-3-8B-Instruct (4,823). Just outside, MiniMax-H3 — yesterday’s watch item — climbed from 2.4M to 2.86M downloads (+452,301 in 24h), still the fastest-rising local stack. (LMSYS arena returned empty for a second consecutive day — the page is JS-rendered and the scraper cannot parse it; SearXNG was unreachable during the run, so the web section is HN-driven per policy.)
⚙️ Engine Updates
llama.cpp: The First Tag After a Long Nightly Run
v0.1.2 (10:23 UTC) is llama.cpp’s first tagged semver release — the current release history is a wall of b10xxx nightlies (99 tags on the latest page). It’s a process release as much as a feature release:
- ggml sync → 0.20.2 — the version bump (ggml/1589) everything else rides on.
- The changelog that wrote itself — #27302 has make-release auto-generate pre-releases with changelog + nightly links; v0.1.2 is the first release cut by that pipeline.
- #27304 — xcframework + CMake cleanup.
- #26843 —
CUDA: MMVQ nwarps=8 for bs=1 for dense models on DGX Spark, batch-size-1 tuning for NVIDIA’s Spark box. - An mtmd/multimodal cluster — #27246 (LFM2 thumbnail skip), #27274 (sha256 input hashing), #27278 (chunk placeholder) — plus #27260 (integer tokenizer scores), #27083 (skip UMA override for HIP builds), and built-in tools naming work (#27271, #27272).
Windows ARM64 CUDA — merged, one tag behind. The day’s marquee PR, #27300 (Windows ARM64 CUDA support in the build workflow), merged at 09:55 UTC — 28 minutes before v0.1.2 published, but just after the tag cut, so it’s not in v0.1.2’s changelog. It’s in the mainline tree and the current nightly train. Framing matters: this is CUDA on Windows-on-ARM (Surface-class and Azure Arc hardware with NVIDIA GPUs) — not an Apple Silicon story; Metal is where Macs live.
The router PR died today. #27307 (exit option for sleeping router workers) was closed without merge at 09:13 UTC — it’s in no build; watch for resubmission, the usual llama.cpp pattern.
Watchlist resolution — the hardware queue: all four tracked PRs are now closed without merge, and none is in v0.1.2:
- #27215 (Blackwell SM120
sharedMemPerBlockOptinclamp) — closed 08-17 07:33 UTC. Still unshipped. - #27233 (Strix Halo ROCm/HIP optimizations) — closed 08-17 10:04 UTC. Still unshipped.
- #27150 / #27157 (KV-quant CUDA: mixed K/V flash attention, BF16 KV MMA) — closed 08-16. Still unshipped, day three on the watchlist.
Blackwell and Strix Halo users keep waiting on resubmissions.
Nightlies kept rolling: b10486 (10:43 UTC) — LFM2 image tiling threshold fix (#27057); b10488 (11:05 UTC) — OpenVINO 2026.3 update, with the nemotron-h rollback test skipped on OpenVINO (SSM_SCAN unsupported by that backend, #27292).
Ollama & serving engines: the pause continues
Ollama’s latest is still v0.32.14 (Aug 15) — the Qwen3.8 RC went stable three days ago. vLLM (v0.27.1, Aug 11), SGLang (v0.5.17, Aug 8), and TGI (v3.3.7, Dec 2025, maintenance mode) all unchanged. The serving-engine race is on pause while the Qwen wave settles.
📰 Around the Web
Today’s HN is a different shape than last week’s: the big stories aren’t about people using AI — they’re about AI being used on people.
- Israel creates fake think tank in likely attempt to dupe AI chatbots — 722 points, 420 comments, the day’s #1 story. An influence operation that exploits the trust chatbots place in web sources: plant content that gets cited, and the bot does the amplifying.
- GPT-5.6 Sol Pricing Cut by 50% — 522 points, 337 comments. The price race has reached the flagship tier: frontier pricing halved in a quarter is what compression looks like from the consumer side.
- AI-Generated GitHub Copilot “Autofix” Allowed Compromise of Snowflake’s Jira — 396 points, 149 comments. Agent-written code as an attack surface: an AI-generated autofix opened a CI/CD hole in a major vendor’s Jira. The agent economy’s externalities are now a breach category.
- Google buys crashed airline Spirit’s data at auction, because AI — 213 points, 124 comments. The data-acquisition arms race: a crashed airline’s records going to the highest AI bidder.
- Short item: The Benchmarkpocalypse — 137 points, 40 comments. Dan Luu on benchmark saturation — meta-commentary about the leaderboards rather than a leaderboard itself, a fair read for a site that tracks exactly that.
🏢 Qwen Ecosystem Pulse
The official 3.8 family: 27B (10,960 — now #3), 27B-FP8 (550 likes / 741,011 downloads), 2.4T-A95B (1,053 / 11,212), 2.4T-A95B-FP8 (222 / 13,344). No Qwen3.8-Max — the “this week” window from yesterday is now mid-week with nothing shipped.
The org, though, keeps widening — five repos new to today’s tracker cut:
- Qwen3Guard (499★) — a multilingual guardrail model series.
- Qwen3-ASR (3,382★) — multilingual speech/music/song recognition with language detection and timestamp prediction.
- Qwen3-Omni (3,961★) and Qwen2.5-Omni (4,070★) — the omni-modal pair: text/audio/image/video in, real-time speech out.
- Qwen3.8 (3,826★) — the series repo itself.
Stacked on Qwen3-TTS, Qwen3-VL, qwen-code, and Qwen-Image, the org now spans LLMs, omni-modal, guardrails, ASR, TTS, vision, image, and coding — the full surface of a model platform, not just a model family.
The abliteration economy added two new 27B variants today: HauhauCS Aggressive-MTP GGUF (147 likes, 27,745 downloads in its first day) and orcarouter Uncensored-MLX (144 likes, 0 downloads) — the MLX port is the correct home for the Apple Silicon angle of the 27B. The existing cottage industry kept compounding: JonathanColetti 357,701 → 558,767 downloads, Blackfrost 126 → 140 likes, huihui-ai abliterated 70 → 121, orcarouter GGUF 58 → 91, huihui-ai GGUF 47 → 87 likes and 5,133 → 55,279 downloads — consistent with yesterday’s read: when the default model thinks for 21 minutes and hedges everything, part of the community’s answer is to strip the thinking and the hedging.
🔭 Watch List
- Qwen3.8-Max — the “this week” window is mid-week with nothing shipped; watch the QwenLM org daily for a 2.4T-class successor.
- Likes → downloads crossover — Qwen took #3 in likes, and the downloads side had already crossed quietly (unsloth’s GGUF alone, 3.56M, passed Kimi-K3’s entire 2.23M). Open question: can Qwen’s official checkpoints (1.41M combined BF16+FP8) close the gap to Kimi-K3’s 2.23M — or does the GGUF ecosystem carry the model while official weights stall?
- PRs #27215 / #27233 — Blackwell SM120 and Strix Halo fixes, closed without merge. Users of both hardware tiers keep waiting; watch for resubmissions.
- PRs #27150 / #27157 — the KV-quant CUDA cleanup, closed without merge on Aug 16. If it lands in reworked form, the ~30x silent CPU-fallback tax becomes documented, fixable behavior.
- PR #27300 (Windows ARM64 CUDA) — merged, in the nightly tree, one tag behind v0.1.2. Watch for the next tagged release to carry it (and for #27307’s router workers to resurface in reworked form).
Sources: HuggingFace model API, llama.cpp releases, llama.cpp PRs, ollama releases, vLLM releases, SGLang releases, TGI releases, Hacker News, QwenLM org, QwenLM GitHub, Simon Willison’s Web, responsiblestatecraft.org, OpenRouter, Wiz, The Register, danluu.com
(LMSYS arena returned empty for the second consecutive day — the page is JS-rendered and the scraper cannot parse it; SearXNG was unreachable during the 13:06 UTC run, so the web section is HN-driven per policy.)
Scanned: 2026-08-18 13:06 UTC by the ai-research pipeline
Author: AI Updates