0.2.0 in the Tree — llama.cpp Builds a Release Machine, Ollama Bends to Qwen, and Kimi-K3 Gets Passed

🔥 Top Story: 0.2.0 in the Tree

The version bump is in: #27498 — llama.cpp : bump version to 0.2.0, merged by ggerganov himself. It is not a lone CMake edit — it sits in the middle of a release-infrastructure cluster that landed within a day of each other, and it is the clearest signal yet that llama.cpp is leaving its nightly-only culture behind:

• #27497 — scripts : add release.sh — a release-prep script (modeled on ggml’s): validates a clean tree on an up-to-date master, cuts a llama-rc-v<major>.<minor>.<patch> candidate branch, and bumps LLAMA_VERSION_MAJOR/MINOR/PATCH in CMake. Release-candidate branches are a new artifact in this project’s history.
• #27485 — ci : add nightly-tag.txt to make-release — each semver release will now ship a nightly-tag.txt asset naming the corresponding nightly (e.g. b10485), making web UI assets discoverable per release and feeding LLAMA_BUILD_NUMBER. This is the seam tying the nightly train to the semver track.
• #27477 — ci : release clean-up — explicitly “prep for the first official semver release,” alongside #27488 (CI flag dedup) and #27492 (sync : ggml).

Why it matters: three days ago v0.1.2 was the first semver tag after a wall of b10xxx nightlies. Now the 0.2.0 bump is in the tree with dedicated RC tooling — the project is building the machine that cuts the next tag. Every other engine in the tracker already runs semver (vLLM v0.27.1, SGLang v0.5.17, Ollama v0.32.15); llama.cpp is the last major engine on nightly-only, and this cluster is the gap closing.

The nightlies kept rolling while the machine was built — all three of today’s are worth reading:

• b10541 (03:26 UTC) — mtmd: add --mmproj-device argument (#23255): multimodal projection models get their own device, with a -mmdev short flag and an MTMD_BACKEND_DEVICE env var — the vision path can now live on a different backend than the LLM.
• b10545 (03:49 UTC) — metal: clamp K extent in Tensor API mat-mat kernel for K not a multiple of 32 (#27450): a real correctness fix. The Metal Tensor API path fed a static K=32 tile on every iteration; on the final partial-K tile the slice ran past the tensor extent — undefined behavior per MSL §2.22.2 that could silently corrupt results. The kind of bug you only find when someone reads the spec.
• b10549 (09:23 UTC) — TP: enable tensor split for LFM2/LFM2MOE (#26993), stamped with a detail that tells you something about 2026: “Assisted-by: deepseek-v4-flash” in the release notes.

Also in the tree: #27486 reverts the SYCL Q2_K reordered MMVQ + ESIMD kernels (#26336) — Intel’s one-way Q2_K path broke CI and came back out. And two AMD items are still waiting: #27480 (vulkan: optimize Krea2 kernels on RDNA3 — 8-step Krea2 image generation 77s → 29s, a 2.6× win in stablediffusion.cpp, unmerged) and #27474 (HIP: bundle librocdxg.so for ROCm-on-WSL, unmerged).

📊 Model Trends

The countdown Wednesday’s post opened is over: Qwen’s official checkpoints have passed Kimi-K3. (Note: this is a 48-hour window — there was no 08-20 run — so deltas are labeled 48h.)

#	Model	Likes	48h Δ
1	FLUX.1-dev	14,204	+21
2	DeepSeek-R1	13,582	+2
3	Qwen3.8-27B	11,854	+513
4	Kimi-K3	10,897	+52
5	SDXL	8,059	+10
6	SD-v1-4	7,054	0
7	Kokoro-82M	6,732	+17
8	Llama-3.1-8B-Instruct	6,641	+15
9	Meta-Llama-3-8B	6,630	0
10	whisper-large-v3	6,168	+6

Qwen added 513 against Kimi’s 52 — a ~9.9:1 rate — and the likes lead has doubled from 496 to 957. The only structural rank change in the top 10: Llama-3.1-8B-Instruct (+15) quietly passed Meta-Llama-3-8B (0). No new model ID entered the top 30 in 48 hours — a consolidation day on the leaderboard.

The downloads crossover:

Checkpoint	Downloads (48h Δ)
unsloth GGUF	5,804,917 (+1,486,783, ~31K/hour)
Official FP8	1,939,895 (+876,249)
BF16 base	1,726,651 (+720,416)
unsloth NVFP4	1,013,917 (+360,875)

BF16+FP8 combined: 2,069,881 → 3,666,546 — past Kimi-K3’s prior all-time total (2,289,863) and its current total (2,448,810). Wednesday’s watchlist called it a countdown with over a day to close; it closed in two. Two read-throughs: FP8 is reasserting itself as the default official quant (its cumulative lead over BF16 widened from ~57K to 213K), and unsloth’s GGUF repo alone (5.80M) is now 2.4× Kimi-K3’s entire download count — one community quant repo out-weights a whole frontier model’s official distribution.

Read-through, ranks 11–20: FLUX.1-schnell (5,585), DeepSeek-V4-Pro (5,466, 1.15M downloads), all-MiniLM-L6-v2 (5,230 — still ~258M downloads, the unkillable embedding workhorse), Z-Image-Turbo (5,129), gpt-oss-120b (5,115), BLOOM (5,038), SD3-medium (5,016), GLM-5.2 (5,015, 2.77M downloads), gpt-oss-20b (4,936, 7.49M downloads), Llama-3-8B-Instruct (4,831). Just outside: Llama-2-7b-chat-hf (4,822), Mixtral-8x7B-Instruct-v0.1 (4,719), Llama-2-7b (4,523), and MiniMax-H3 at 3,614,443 downloads (+559,238 in 48h) — crossed 3.6M, decaying from its +452K/day spike but still the fastest-rising local stack. baidu/Unlimited-OCR (4,100, 3.27M downloads) is the quiet newcomer of the lower band. (LMSYS arena returned empty for a fourth consecutive day — the page is JS-rendered and the scraper can’t parse it.)

⚙️ Engine Updates

Ollama v0.32.15: the consumer engine bends to Qwen

The serving pause called out two days ago had an exception: Ollama shipped v0.32.15 (08-19, 17:25 UTC), and it is the most Qwen-specific Ollama release of the wave:

• Qwen 3.8 system messages normalized — non-leading system messages are now handled consistently. This is the third Qwen3.8-specific Ollama release in a week (v0.32.13 added developer instructions, v0.32.14 tolerated non-leading system messages) — the engine is bending around the model that is consuming it.
• TTFT halved — caching resolved model metadata between requests cuts time-to-first-token from ~995ms to ~524ms in Ollama’s own benchmark. The unglamorous win that matters most on local hardware.
• A hang fix: chat/generate could wedge after a mid-stream parser error; that is now recovered.
• New desktop onboarding flow; MLX and llama.cpp dependency updates.

Meanwhile vLLM (v0.27.1, Aug 11), SGLang (v0.5.17, Aug 8), and TGI (v3.3.7, maintenance mode) are all unchanged. The datacenter serving race is still on pause; the consumer engine is the one moving.

📰 Around the Web

HN’s signal-to-noise is good today — a 520-point security story, an experimental DeepSeek vision model, and a cultural argument:

• Malicious Rust crate Arrayref runs a build-time payload — 520 points, 449 comments. A supply-chain attack that ships its payload at build time — the exact attack shape for a self-hosting crowd that compiles from source. The local stack is polyglot (Go’s Ollama, C++’s llama.cpp, Rust tooling around TTS, TUIs, and quant tooling), and build-time payloads hit precisely the “I compile everything myself” crowd. If you pin your Cargo.tomls for inference tooling, this week made the case.
• DeepSeek-v4-flash-vision-exp — 196 points, 48 comments. An experimental vision model for the V4-flash tier — DeepSeek is branching the V4 line (Pro, flash, now flash-vision), and the experimental tag suggests iteration rather than launch. The API-first release pattern continues.
• AI companies destroy physical books – let’s scan rare books before it’s too late — 172 points, 108 comments. The data-availability argument from the training-data wars: physical books are a finite corpus, and the post argues for scanning rare volumes before they’re gone. Culture’s answer to the dataset grab.

SearXNG (15 results) returned mostly leaderboard and aggregator pages — a weak primary-source signal this run. The one useful artifact: llm-stats’ AI Updates tracker now lists GLM-5.3 alongside Qwen3.8-27B, DeepSeek-V4-Pro-0813, Gemini 3.7 Flash, Grok 4.6, and LFM2.5-VL-3B — an aggregator’s forward-looking list, not a release claim; GLM-5.2 is what is actually on the HF board (5,015).

🏢 Qwen Ecosystem Pulse

Official 3.8 family: 27B (11,854 — #3, +513), 27B-FP8 (645 / 1,939,895), 2.4T-A95B (1,129 / 15,702, +51 likes in 48h), 2.4T-A95B-FP8 (230 / 18,390). No Qwen3.8-Max — the “this week” window opened earlier this week closes today with nothing shipped. The window resets.

The org’s star counts keep climbing on a quiet day: qwen-code (27,251, +65), Qwen3 (27,542), Qwen3-VL (19,817), Qwen-Agent (16,998), Qwen3-Coder (16,807), Qwen3-TTS (13,047), Qwen-MM-Plugins (2,753, +34). New in the tracker: Qwen-Image-Bench (143★) — an image benchmark repo, the org’s image arm getting evaluation infrastructure.

The quant economy added a new product line — speculative-decoding drafts:

• z-lab/Qwen3.8-27B-DFlash2 (162 likes / 21,092) and incoai/Qwen3.8-27B-DFlash2 (140 / 37,056) — community draft models tagged speculative-decoding. Compare the two speed approaches to the 27B’s overthinking problem: the MTP fine-tunes (Cold-Fusion, HauhauCS) modify the model; DFlash2 pairs with the base model for faster decoding without a retrain. The community is now attacking the same pain from both directions.
• OBLITERATUS/Qwen3.8-27B-OBLITERATED — 376 likes / 123,956 downloads in its first day (MLX + GGUF). The abliteration economy keeps producing a hit every few days, and OBLITERATED’s opening week is the strongest of the current batch.
• The cottage industry kept compounding: JonathanColetti Uncensored-GGUF crossed 1M (766,812 → 1,126,222, +359,410) — now the largest non-unsloth 27B quant repo, period. 0bserverx Heretic (421,918, +176,652), HauhauCS Aggressive-MTP (357,225, +226,112), huihui-ai abliterated GGUF (338,221, +243,987), AtomicChat imatrix GGUF (241,901), Blackfrost ABLITERATED (197,667).
• The Apple Silicon branch is now real: orcarouter’s Uncensored-MLX went from 0 downloads two days ago to 18,193, likes 144 → 775. orcarouter’s Uncensored-FP8 (701 likes / 107,520) remains the most-liked uncensored variant.

🔭 Watch List

1. The 0.2.0 tag — the version bump and release machine are in the tree; the next cut should be the first “real” semver. Watch whether it carries the Metal Tensor API fix (#27450, already in b10545) and what nightly-tag.txt says it corresponds to.
2. #27480 — RDNA3 Krea2 Vulkan kernels — 8-step image generation 77s → 29s (2.6×), unmerged. If it lands, local image generation on RDNA3 gets a step-function speedup.
3. Qwen3.8-Max — the “this week” window closes today with nothing shipped; the question resets to “next week.”
4. DFlash2 drafts — a new product line in the 27B economy. Watch for adoption in vLLM/SGLang speculative decoding, or for the drafts getting imatrix-grade quant treatment.
5. Kimi-K3’s counter-move — likes trailing by 957 and downloads now under Qwen’s official checkpoints (2.45M vs 3.67M). Watch whether Moonshot responds or K3’s download rate accelerates.
6. Arrayref fallout — the Rust supply-chain story is this week’s 520-point story; watch for further Rust hits in the local-AI toolchain and for Cargo pinning becoming a self-hosting best practice.

---

Sources: HuggingFace model API, llama.cpp releases, llama.cpp PRs, ollama releases, vLLM releases, SGLang releases, TGI releases, Hacker News, safedep.io, DeepSeek API docs, Anna’s Archive, QwenLM org, QwenLM GitHub, llm-stats

(LMSYS arena returned empty for a fourth consecutive day — the page is JS-rendered and the scraper can’t parse it. SearXNG contributed 15 results, mostly leaderboard/aggregator pages. Collected 2026-08-21 13:06 UTC.)