The Archive

Discrete diffusion language model (DiffusionGemma 26B MOE) transcribes speech in parallel via denoising instead of autoregressive decoding.

DeepReinforce releases Ornith-1.0, MIT-licensed open-weights model (9B–397B variants) for agentic coding, built on Gemma 4 and Qwen 3.5, achieving SOTA on coding benchmarks.

Three-method study across Qwen2.5-Coder-32B, Llama-3.1-8B, and Gemma-3-27B shows internal probes read situation not pre-action intent, limiting misalignment monitoring efficacy.

Gemma 4 is now significantly faster in Ollama 0.31 on Apple Silicon via multi-token prediction (MTP), powered by MLX. Performance is now up to 90% faster when used with coding agents, as measured using the Aider polyglot benchmark.

LLM reasoning transparency is a critical affordance for understanding model decisions, mitigating misuse and misalignment, and debugging surprising model behaviors. However, DiffusionGemma performs a larger fraction of its computation in a continuous latent space; does this make its reasoning less transparent? We study this question by decomposing transparency into two components: variable transparency, whether we understand intermediate snapshots of a model's computational state; and algorithmic transparency, whether we can use these snapshots to reconstruct the process by which the model ar...

Fine-tuning language models on insecure code induces emergent misalignment with poorly understood internal structure. We investigate whether this misalignment corresponds to a causally actionable activation-space direction shared across architectures. Across four instruction-tuned model families (Qwen2.5-1.5B, Gemma-2-2B, Llama-3.2-1B, Ministral-3-3B) finetuned identically, a difference-in-means direction achieves 99.6% separation of aligned and misaligned activations at each model's final layer. Causal steering by subtracting this direction reduces code spillover by 21-51 points, while a sec...

Open diffusion language models are marketed as parallel, non-autoregressive decoders, yet the order in which a shipped checkpoint actually commits its tokens is almost never measured. We instrument DiffusionGemma 26B, a masked discrete-diffusion mixture-of-experts model built on Gemma 4, hooking its sampler's accept step to record which canvas positions commit, when, and at what confidence. Across a 686-prompt, six-regime probe suite we find that its decoding is neither parallel nor block-autoregressive: it follows a partial left-to-right commit bias whose apparent strength depends almost ent...

Fine-tuning vision-language models to emit dense coordinate lists improves visual grounding but also changes how models serialize, repeat, and terminate structured outputs. We study this behavior as a generation and control surface. In Gemma 4 12B, high-capacity q/k/v/o LoRA raises class-aware F1@0.3 from 0.007 to 0.448 while inducing repeated-tail pressure (duplicate rate 0.080, max repeat 23). A q/v rank sweep keeps max repeat at 21-22 across ranks 4-64, showing capacity persistence. The target signal is separable: object-level repeat-stop removes exact repeated records (duplicate rate 0.00...

DiffusionGemma Last May Google briefly released an experimental Gemini Diffusion model. I tried the preview at the time and recorded it running at 857 tokens/second. It was an exciting model, but Google made no further announcements about it. That research has returned in the best possible way: as a new open weight (Apache 2 licensed) Gemma model, google/diffusiongemma-26B-A4B-it . NVIDIA are currently hosting the model for free on their NIM cloud API. I used that API to generate this pelican , which took 4.4s (according to time uv run generate.py ) to return 2,409 tokens - so at least 500 to...

Diffusion AI is most common in image generation, but it can make text outputs much faster.

Developers building real-time AI—such as chat assistants, copilots, and agentic workflows—are often constrained by token-by-token generation speed. This... Developers building real-time AI—such as chat assistants, copilots, and agentic workflows—are often constrained by token-by-token generation speed. This limits responsiveness, increases serving costs, and makes fluid, interactive experiences difficult to achieve. DiffusionGemma, created by Google DeepMind and optimized to run efficiently across NVIDIA platforms, introduces a new approach to… Source

Gemma 4 12B uses a new encoding scheme and token prediction to punch above its weight.

We describe AlignAtt4LLM, an IWSLT 2026 simultaneous speech translation system for English to German, Italian, and Chinese. The system is a synchronous cascade: Qwen3-ASR with forced alignment produces an incrementally updated source transcript, and Gemma-4 E4B-it translates that prefix under an MT-side AlignAtt policy. To our knowledge, this is the first application of AlignAtt to a decoder-only LLM, where the encoder-decoder cross-attention used by earlier AlignAtt systems is absent. We recover a usable policy by proposing (1) an explicit source span in the prompt, (2) offline selection of ...

Provided in both Safetensors and GGUFs. Safetensors, llmfan46/Gemma-4-Harmonia-31B-it-uncensored-heretic: [https://huggingface.co/llmfan46/Gemma-4-Harmonia-31B-uncensored-heretic](https://huggingface.co/llmfan46/Gemma-4-Harmonia-31B-uncensored-heretic) GGUFs, llmfan46/Gemma-4-Harmonia-31B-it-uncensored-heretic-GGUF: [https://huggingface.co/llmfan46/Gemma-4-Harmonia-31B-uncensored-heretic-GGUF](https://huggingface.co/llmfan46/Gemma-4-Harmonia-31B-uncensored-heretic-GGUF) Comes with benchmark too. Find all my models here: [HuggingFace-LLMFan46](https://huggingface.co/llmfan46/models) The o...

Last month NVIDIA released [SOL-ExecBench](https://research.nvidia.com/benchmarks/sol-execbench), a new benchmark of 235 production CUDA kernels lifted from DeepSeek, Qwen, Gemma, and Kimi. We took several top-ranked AI-generated submissions and tried using them in production workloads. Many of them broke, sometimes in surprising ways. One of those kernels is the fused embedding-gradient + RMSNorm backward pass, which runs at the end of every transformer training step. We took the fastest submission on the benchmark for it, and dropped it into the training loop of a small transformer. The ke...

I don't see any threads on this model. Is it because it's dense and/or without-**reasoning**? Anyone tried this for coding? >[**Capabilities**](https://huggingface.co/ibm-granite/granite-4.1-30b) Summarization Text classification Text extraction Question-answering Retrieval Augmented Generation (RAG) Code related tasks Function-calling tasks Multilingual dialog use cases Fill-In-the-Middle (FIM) code completions Some people prefer dense in this model size range(Ex: 27B over 35B-A3B). Still no feedbacks from them here. I know that some people love Granite models. Myself...

User reports Qwen3.6 35B outperforms Gemma4, GLM 4.7 Flash, others for local agentic tasks; seeks comparable MoE alternatives.

Reddit discussion comparing inference speed/quality tradeoffs between Qwen3.6-35B and Gemma4-26B on consumer GPU hardware.

Chrome extension enables local inference of Gemini Nano (Gemma) on CPU-only systems, ~20 tokens/sec on laptop.

Wanted to share a workflow I tested on a real flight, in case anyone else is trying to set up offline Claude Code. The core idea: using ollama to pull the needed model of what you need, and then use it to run claude code The setup, in order: 1. Pull a model on home wifi the night before. \`ollama pull <model>\` — \~9 GB for a 14B, \~17 GB for a 26B. Don't try this at the gate. 2. In Claude Code, point at Ollama. The cleanest path I found is wrapping it in two aliases: alias claude-local='ollama launch claude --model gemma4:26b' alias claude-cloud='claude' 3. Verify on the ground with...

User reports Apex quantization for Gemma 4 26B achieves 38 tokens/sec at 90K context on RX 9060 XT with llama.cpp.

Community finetune of Gemma 4 26B with reduced refusals; niche interest for local model enthusiasts.

BeeLlama v0.2.0 achieves 4-5x token throughput gains on RTX 3090 via DFlash optimizations for Qwen 27B and Gemma 31B models.

LatitudeGames releases Equinox-31B, a Gemma-based 31B finetune blending dark adventure and slice-of-life storytelling data.

Early-stage Gemma 4 MTP compilation work-in-progress shared on LocalLLaMA.

Google AI Edge Gallery v1.0.13–14 adds Gemma 4 multi-token prediction, Pixel TPU support, experimental MCP, and chat history persistence.

Trace-based benchmark measuring safety-aligned LLM behavior (Gemma 4) as autonomous security agents vs. uncensored derivatives on 30 vulnerability-analysis tasks.