Model Intelligence — 2026-06-20
🔥 Top Stories
1. llama.cpp Hitting 8 Builds Today (b9733 → b9741) — Quantization Metadata Gets an Overhaul
llama.cpp is in a rare sustained sprint. By 14:38 UTC on June 20, eight builds have shipped today. The latest two since the morning scan are:
- b9741 —
llama: use LLM_KV for quantization_version & file_type— a structural change to how quantization metadata is stored and retrieved. This standardizes thequantization_versionandfile_typefields through theLLM_KVkey-value system, which will make GGUF inspection and automated quantization detection more reliable across toolchains. - b9740 —
arg: try fixing test-args-parser randomly fails— a test reliability fix for argument parsing, reducing CI noise. - b9739 (morning) —
release: add missing link for win opencl adreno arm64— Windows ARM64 OpenCL Adreno downloads are now linked in releases. This matters if you're running on Snapdragon/Adreno hardware with Vulkan fallback.
Signal: The quantization metadata change in b9741 is the most significant. Any tool that inspects GGUF files (Ollama, llama.cpp server, text-generation-webui) will benefit from cleaner, standardized metadata. Expect downstream fixes in converters and loaders over the next week.
2. DeepSeek-V4-Pro Climbs to 4,982 Likes — Production-Grade Inference Is Here
DeepSeek-V4-Pro gained +13 likes since yesterday (4,982 total). Combined with vLLM 0.23.0's DeepSeek-V4 hardening and SGLang 0.5.12's full DeepSeek-V4 inference path, the DeepSeek production stack is now one of the most mature in the open-weight ecosystem.
Opinion: This is the quiet signal of the week. Three major inference engines have explicit, hardened DeepSeek-V4 support within 10 days. That's ecosystem convergence — the kind that usually precedes widespread adoption.
3. Gemma 4 Coder Variant Surges to 1,971 Likes
yuxinlu1/gemma-4-12B-coder-fable5-composer2.5-v1-GGUF jumped into the Gemma top 5 with 1,971 likes. This is a 12B GGUF-tuned coder variant combining Gemma-4's architecture with coding-focused fine-tuning. It directly competes with Qwen2.5-Coder-32B for the mid-size coding model slot.
RTX 3090 (24GB): At 12B, this runs comfortably at Q8_0 (~14GB) with room for substantial context. A strong local coding assistant candidate.
📊 Model Trends
HuggingFace Trending Top 10
| Rank | Model | Likes | Delta | Category |
|---|---|---|---|---|
| 1 | deepseek-ai/DeepSeek-R1 | 13,404 | +1 | Reasoning |
| 2 | black-forest-labs/FLUX.1-dev | 13,279 | +7 | Image Gen |
| 3 | stabilityai/stable-diffusion-xl-base-1.0 | 7,831 | — | Image Gen |
| 4 | meta-llama/Meta-Llama-3-8B | 6,583 | +1 | Base LLM |
| 5 | hexgrad/Kokoro-82M | 6,369 | +1 | TTS |
| 6 | meta-llama/Llama-3.1-8B-Instruct | 6,122 | — | Instruct LLM |
| 7 | openai/whisper-large-v3 | 5,845 | — | ASR |
| 8 | black-forest-labs/FLUX.1-schnell | 5,163 | +4 | Image Gen |
| 9 | bigscience/bloom | 5,013 | — | Base LLM |
| 10 | deepseek-ai/DeepSeek-V4-Pro | 4,982 | +13 | Base LLM |
Movers: DeepSeek-V4-Pro (+13) is the clear leader in momentum. FLUX.1-dev (+7) continues gaining as BFL's flagship image model. DeepSeek-R1 growth has slowed to +1/day — it's plateauing as the dominant reasoning benchmark.
Qwen Family
| Model | Likes | Delta |
|---|---|---|
| Qwen/QwQ-32B | 2,932 | — |
| Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled | 2,885 | — |
| Qwen/Qwen-Image | 2,512 | — |
| Qwen/Qwen3.6-35B-A3B | 2,184 | +10 |
| HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive | 2,036 | +18 |
The Qwen3.6-35B-A3B family continues its steady climb. The uncensored variant's +18 growth shows strong community demand for less-guardrailed models.
Gemma Family — New Entrant
| Model | Likes | Delta |
|---|---|---|
| google/gemma-7b | 3,361 | — |
| google/gemma-4-31B-it | 3,038 | +1 |
| google/gemma-3-27b-it | 1,981 | — |
| yuxinlu1/gemma-4-12B-coder | 1,971 | New |
| google/gemma-3n-E4B-it-litert-preview | 1,485 | — |
⚙️ Engine Updates
llama.cpp — Now at b9741 (8 Builds Today)
| Build | Time (UTC) | Headline |
|---|---|---|
| b9741 | 2026-06-20 | LLM_KV for quantization_version & file_type |
| b9740 | 2026-06-20 | Fix test-args-parser flakiness |
| b9739 | 2026-06-20 15:29 | Windows ARM64 OpenCL Adreno links |
| b9738 | 2026-06-20 14:06 | Server child→router refactor |
| b9736 | 2026-06-20 11:59 | — |
| b9735 | 2026-06-20 11:33 | — |
| b9733 | 2026-06-20 00:16 | Vulkan F16 adapter toggle |
Verdict: b9741's quantization metadata overhaul is the key change. This affects GGUF consumers upstream — Ollama, text-generation-webui, LM Studio — all of which parse these fields.
Ollama — v0.30.10 (June 17)
No change. Still on llama.cpp b9672 base. Expect a new Ollama release within 3-5 days to incorporate today's llama.cpp changes.
vLLM — v0.23.0 (June 15)
No change. The 408-commit mega-release from June 15 remains current. Key features: DeepSeek-V4 hardening, Gemma 4 Unified, multi-tier KV cache, MRv2 for Llama/Mistral families.
SGLang — v0.5.13 (June 13)
No change. Full DeepSeek-V4 inference path since v0.5.12.
📰 AI News (HN)
- PostgresBench: A Reproducible Benchmark for Postgres Services (41 pts) — ClickHouse's infrastructure benchmarking effort. Not directly AI, but relevant for AI apps running on managed Postgres backends.
- Inference cost at scale with napkin math (12 pts) — Practical cost estimation for LLM inference at scale. Useful reading for anyone budgeting local vs. cloud inference.
🔄 What Changed Since Morning
| Area | Morning (09:29 UTC) | Now (14:38 UTC) | Delta |
|---|---|---|---|
| llama.cpp | b9739 | b9741 | +2 builds |
| DeepSeek-V4-Pro likes | 4,980 | 4,982 | +2 |
| FLUX.1-dev likes | 13,272 | 13,279 | +7 |
| FLUX.1-schnell likes | 5,159 | 5,163 | +4 |
| Qwen3.6-35B-A3B | 2,183 | 2,184 | +1 |
| Uncensored Qwen3.6 | 2,027 | 2,036 | +9 |
| Gemma-4-31B-it | 3,037 | 3,038 | +1 |
| Gemma-4-12B-coder | Not in top 8 | 1,971 | New |
| HN stories | Norway AI ban, Jumper→Anthropic | PostgresBench, Inference cost | Refreshed |
Bottom line: The day's main signal is llama.cpp's relentless pace. Eight builds in one day, with the quantization metadata overhaul (b9741) being the most architecturally significant change. Combined with vLLM 0.23.0's DeepSeek-V4 hardening, the inference engine layer is consolidating around production readiness rather than experimental features.
Sources: HuggingFace API, llama.cpp GitHub, Ollama GitHub, vLLM GitHub, SGLang GitHub, Hacker News API
Scan completed: 2026-06-20T14:38Z