Model Intelligence — 2026-06-20
llama.cpp surges with 6 builds today; DeepSeek-R1 climbs 13,403 likes; vLLM 0.23.0 brings major DeepSeek-V4 hardening and Gemma 4 Unified support.
🔥 Top Stories
1. llama.cpp Goes Full Speed on June 20 — Six New Builds in One Day
The llama.cpp project is burning hot. Between 00:16 UTC and 15:29 UTC on June 20, the ggml-org team pushed six builds (b9733 through b9739), including a Vulkan F16 adapter toggle for NVIDIA and a major refactor of child-to-router communication in the server layer.
What this means for you: The Vulkan F16 adapter toggle (b9739) directly impacts GPU inference on non-CUDA setups. If you’re running on AMD GPUs with Vulkan, you can now force F16 precision — a significant quality-of-life improvement that could mean fewer precision-related artifacts in text generation. The server-side refactor (b9738) optimizes token probability lookups, which translates to faster sampling in multi-threaded serving scenarios.
RTX 3060 (12GB): You can run Llama-3.1-8B-Instruct at Q4_K_M (~5.3GB VRAM) comfortably. The new Vulkan adapter helps if you’re using AMD. For Q8_0 (~10.2GB), you’ll be tight on 12GB but feasible with quantization offloading.
RTX 3090 (24GB): You can load Qwen-Image-Edit or QwQ-32B at Q4_K_M (~18-20GB VRAM). The 32B-class models are where llama.cpp really shines on 24GB cards.
2. DeepSeek-R1 Dominates the Leaderboard at 13,403 Likes
DeepSeek-R1 continues its reign atop HuggingFace’s trending models with 13,403 likes (+2 since last scan). The reasoning-focused model from DeepSeek has become the de facto benchmark for open-weight reasoning capabilities.
The ecosystem effect: The rise of DeepSeek-V4-Pro (4,980 likes, +11 since yesterday) signals a maturing DeepSeek family. Combined with vLLM 0.23.0’s explicit DeepSeek-V4 hardening, production deployment of these models is becoming far more stable. If you’ve been hesitating to try DeepSeek models in your pipeline, the inference tooling is now solid.
RTX 3060 (12GB): DeepSeek-R1’s smaller variants (1.5B) fit comfortably. The full model is too large for consumer hardware — look at distilled Qwen variants (QwQ-32B at Q4) if you want reasoning-style outputs in the 20GB range.
RTX 3090 (24GB): QwQ-32B (Qwen’s reasoning model) at Q4_K_M uses roughly 19-20GB — right at the edge but loadable. The Jackrong distilled version (2,885 likes) is another strong option for reasoning-heavy tasks.
3. vLLM 0.23.0: A Massive Release with 408 Commits
vLLM v0.23.0 dropped on June 15 and is the most significant inference engine release in months. Key features:
- DeepSeek-V4 hardening — production-grade support for the latest DeepSeek architecture
- MRv2 (Multi-Resolution v2) for Llama and Mistral families — improved KV cache efficiency
- Gemma 4 Unified support — native handling of Google’s latest Gemma variant
- Multi-tier KV cache — better memory management for long-context workloads
- Rust frontend (experimental) — early access to a Rust-based serving layer
Why it matters: The multi-tier KV cache is a game-changer for long-context local inference. If you’re running models with 128K+ context windows, this reduces memory pressure significantly — meaning you can fit larger context windows on the same hardware.
📊 Model Trends
HuggingFace Trending Top 15
| Rank | Model | Likes | Delta | Key Tags |
|---|---|---|---|---|
| 1 | deepseek-ai/DeepSeek-R1 | 13,403 | +2 | transformers, safetensors, deepseek_v3 |
| 2 | black-forest-labs/FLUX.1-dev | 13,272 | +6 | diffusers, safetensors, text-to-image |
| 3 | stabilityai/stable-diffusion-xl-base-1.0 | 7,831 | +2 | diffusers, onnx, safetensors |
| 4 | CompVis/stable-diffusion-v1-4 | 7,024 | +2 | diffusers, safetensors, stable-diffusion |
| 5 | meta-llama/Meta-Llama-3-8B | 6,582 | +3 | transformers, safetensors, llama |
| 6 | hexgrad/Kokoro-82M | 6,368 | +3 | text-to-speech, en |
| 7 | meta-llama/Llama-3.1-8B-Instruct | 6,122 | +5 | transformers, safetensors, llama |
| 8 | openai/whisper-large-v3 | 5,845 | +7 | transformers, pytorch, jax |
| 9 | black-forest-labs/FLUX.1-schnell | 5,159 | +5 | diffusers, safetensors, text-to-image |
| 10 | bigscience/bloom | 5,013 | +1 | transformers, pytorch, tensorboard |
| 11 | deepseek-ai/DeepSeek-V4-Pro | 4,980 | +11 | transformers, safetensors, deepseek_v4 |
| 12 | sentence-transformers/all-MiniLM-L6-v2 | 4,978 | +2 | sentence-transformers, pytorch, tf |
| 13 | stabilityai/stable-diffusion-3-medium | 4,977 | +1 | diffusion-single-file, text-to-image |
| 14 | openai/gpt-oss-120b | 4,905 | +6 | transformers, safetensors, gpt_oss |
| 15 | Tongyi-MAI/Z-Image-Turbo | 4,839 | +2 | diffusers, safetensors, text-to-image |
Notable shifts: DeepSeek-V4-Pro (+11 likes) is the biggest mover in the trending list. The openai/gpt-oss-120b entry is new in the top 15, suggesting growing interest in OpenAI’s open-source 120B parameter model. Tongyi-MAI’s Z-Image-Turbo also appeared new, reflecting Alibaba’s push into open image generation.
Qwen Model Family
| Model | Likes | Delta |
|---|---|---|
| Qwen/QwQ-32B | 2,932 | +1 |
| Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled | 2,885 | 0 |
| Qwen/Qwen-Image | 2,512 | 0 |
| Qwen/Qwen-Image-Edit | 2,426 | +1 |
| Qwen/Qwen3.6-35B-A3B | 2,183 | +9 |
| Phr00t/Qwen-Image-Edit-Rapid-AIO | 2,171 | +1 |
| Qwen/Qwen2.5-Coder-32B-Instruct | 2,046 | 0 |
| HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive | 2,027 | +17 |
| Qwen/Qwen2.5-Omni-7B | 1,909 | 0 |
| Qwen/Qwen3.6-27B | 1,762 | +6 |
Qwen analysis: The Qwen3.6-35B-A3B variant is gaining traction rapidly (+9 today). The uncensored variant by HauhauCS jumped +17, suggesting community demand for unfiltered Qwen3.6 models. QwQ-32B remains the top Qwen model overall for reasoning tasks.
RTX 3060 (12GB): Qwen2.5-Omni-7B at Q4_K_M (~4.2GB) or Q8_0 (~8.5GB). The 7B models are your sweet spot.
RTX 3090 (24GB): Qwen3.6-27B at Q4_K_M (~15GB) or Q6_K (~18GB) — plenty of room for context. QwQ-32B at Q4_K_M (~19GB) fits with tight margin.
Google Gemma Family
| Model | Likes | Delta |
|---|---|---|
| google/gemma-7b | 3,361 | +2 |
| google/gemma-4-31B-it | 3,037 | +5 |
| google/gemma-3-27b-it | 1,981 | 0 |
| google/gemma-3n-E4B-it-litert-preview | 1,485 | 0 |
| google/gemma-2-2b-it | 1,396 | 0 |
| google/gemma-3-4b-it | 1,372 | 0 |
| google/gemma-4-E4B-it | 1,267 | +3 |
| google/gemma-7b-it | 1,248 | +1 |
| google/gemma-2b | 1,196 | +1 |
| google/gemma-4-26B-A4B-it | 1,164 | +2 |
Gemma analysis: Gemma-4-31B-it is the second-most-liked Gemma model and continues growing (+5). The Gemma-4 family (31B and 26B-A4B) is clearly the focus for Google right now. The Gemma-4-E4B-it small model gained +3, showing interest in the efficient 4B variant.
RTX 3060 (12GB): Gemma-7B at Q4_K_M (~4.5GB) or Q8_0 (~9GB). Gemma-4-E4B-it at Q8_0 (~5GB). Both fit comfortably.
RTX 3090 (24GB): Gemma-4-31B-it at Q4_K_M (~19GB) fits perfectly. Q6_K (~24GB) is borderline. The vLLM 0.23.0 Gemma 4 Unified support makes running these smoother than ever.
⚙️ Engine Updates
llama.cpp — 6 New Builds Today (b9733 → b9739)
| Build | Date | Notes |
|---|---|---|
| b9739 | 2026-06-20 15:29 | ggml-webgpu: add adapter toggles for F16 on Vulkan + NVIDIA |
| b9738 | 2026-06-20 14:06 | server: refactor child → router communication |
| b9736 | 2026-06-20 11:59 | — |
| b9735 | 2026-06-20 11:33 | — |
| b9733 | 2026-06-20 00:16 | ggml-webgpu: add adapter toggles for F16 on Vulkan + NVIDIA |
Previous day ended at b9733. Today’s growth: +6 builds. This is an unusually active development day for llama.cpp. The Vulkan F16 support is the headline feature — previously, Vulkan backends had limited precision options, and this toggle gives users direct control.
Source: ggml-org/llama.cpp releases
Ollama — v0.30.10 (June 17)
The latest Ollama release remains v0.30.10, released on June 17. Key features from recent releases:
- v0.30.10: Command A/North family support on Apple Silicon MLX, llama.cpp b9672 base
- v0.30.9: Cohere2Moe support, LFM2 parser fixes (June 15)
- v0.30.8: Prompt caching, MLX stability improvements, recurrent model support (June 12)
No new release today. The next update is expected within days given the rapid llama.cpp development cycle.
Source: ollama/ollama releases
vLLM — v0.23.0 (June 15)
vLLM 0.23.0 remains the latest release, a massive drop with 408 commits from 200 contributors:
- DeepSeek-V4 hardening — production-grade inference for DeepSeek’s latest architecture
- MRv2 (Multi-Resolution v2) — default for Llama and Mistral families, improved KV cache efficiency
- Gemma 4 Unified — native support for Google’s Gemma 4 variants
- Multi-tier KV cache — reduces memory pressure for long-context workloads
- Rust frontend (experimental) — early access to a Rust-based serving layer
- Multi-node support improvements — better distributed inference
Source: vllm-project/vllm releases
SGLang — v0.5.13 (June 13)
SGLang’s latest release remains v0.5.13 from June 13. No new release today. SGLang continues to be a strong alternative for structured generation workflows.
Source: sgl-project/sglang releases
📰 AI News (HN)
Today’s Hacker News front page didn’t feature AI-specific stories prominently. The notable AI-adjacent items from recent scans:
- Norway imposes near ban on AI in elementary schools (514 pts) — regulatory signal from Europe
- John Jumper to join Anthropic (88 pts) — major talent movement in AI research
- Big Tech borrowing like never before (18 pts) — financial pressure on AI infrastructure spending
The CSSQuake project (285 pts) and encryption export control history (76 pts) dominated the front page today.
🔄 What Changed Since Yesterday
New Entries in Trending
- openai/gpt-oss-120b (4,905 likes) — new to top 15, OpenAI’s open-weight 120B parameter model
- Tongyi-MAI/Z-Image-Turbo (4,839 likes) — new to top 15, Alibaba’s image generation model
llama.cpp Activity
- 6 new builds today (b9733 → b9739), up from yesterday’s end at b9733
- Vulkan F16 adapter toggle is the headline feature
- Server communication refactor for better multi-threaded serving
Model Growth
- DeepSeek-V4-Pro +11 likes (biggest mover in trending)
- HauhauCS/Qwen3.6-35B-A3B-Uncensored +17 likes (biggest mover in Qwen family)
- Qwen3.6-35B-A3B +9 likes
- Qwen3.6-27B +6 likes
- openai/whisper-large-v3 +7 likes
- black-forest-labs/FLUX.1-dev +6 likes
- Qwen/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive +17 likes
Stable Models
- DeepSeek-R1, FLUX.1-schnell, and most Gemma variants showed minimal changes (0-3 likes)
- No new major model releases from Meta, Google, or Anthropic detected
Sources: HuggingFace API, llama.cpp GitHub, Ollama GitHub, vLLM GitHub, SGLang GitHub, Hacker News API
Scan completed: 2026-06-20T15:30Z