Releases·Model Release
IFM releases K2-Horizon-7B, diffusion-augmented LLM claiming up to 5200 tps lossless
IFM released K2-Horizon-7B on Hugging Face, a diffusion-augmented LLM that adds a plug-and-play diffusion adapter to a causal autoregressive architecture. IFM claims speeds up to 5200 tokens per second with no quality loss, with a paper on arXiv (2609.04010).
105 engagement1 sourceThu, Sep 17, 2026, 06:43 PM
AI Overview
- IFM released K2-Horizon-7B, a diffusion-augmented LLM claiming up to 5200 tokens/sec with no quality loss
- The model keeps a causal LLM architecture and adds a plug-and-play diffusion adapter alongside the autoregressive weights
- Details are documented on Hugging Face (IFM/K2-Horizon-7B-Uno) and in an arXiv paper (arXiv:2609.04010)
Entities
Related
Releases·Model Release
IFM releases GGUF quants for K2-Horizon models, sizes 0.9B to 36B-A4B MoVA, on Hugging Face
Community-shared GGUF quantized weights for the K2-Horizon model family are now downloadable from Hugging Face under the IFM organization, spanning sizes from 0.9B to a 36B-A4B MoVA variant. Local inference users need a llama.cpp fork for now, since official llama.cpp support remains an in-progress pull request.
Community-shared GGUF quantized weights for the K2-Horizon model family are now downloadable from Hugging Face under the IFM organization, spanning sizes from 0.9B to a 36B-A4B MoVA variant. Local inference users need a llama.cpp fork for now, since official llama.cpp support remains an in-progress pull request.
50 engagement1 sourceredditTue, Sep 22, 2026, 08:26 PM
Developments·Paperfeatured
[Research] ORL hits 2,788 tokens/s on one RTX 5090 with a diffusion LLM — open source
The Introspective Diffusion team introduced ORL (On-Policy Residual Learning), a training method that lets diffusion language models finalize more tokens per forward pass. Combined with their custom CUDA-graph inference engine dlmrt, it reaches 2,788 tokens/s on a single RTX 5090 and up to 6.9× end-to-end throughput gains. Models, training code, and the inference engine are all open source.
The Introspective Diffusion team introduced ORL (On-Policy Residual Learning), a training method that lets diffusion language models finalize more tokens per forward pass. Combined with their custom CUDA-graph inference engine dlmrt, it reaches 2,788 tokens/s on a single RTX 5090 and up to 6.9× end-to-end throughput gains. Models, training code, and the inference engine are all open source.
195 engagement1 sourceredditFri, Oct 16, 2026, 07:33 PM
Developments·Paper
Magenta: Closing the Loop Between Mathematical Reasoning and Lean Verification
A new arXiv paper introduces Magenta, a training-free agentic pipeline that converts natural-language math problems into Lean 4 statements and machine-checked proofs. It reports 100% accuracy on AIME 2025, AIME 2026, and HMMT February 2026, and solves all six IMO 2026 problems when paired with the K2-Horizon-7B reasoner. The work matters to practitioners because it shows formal verification signals can be integrated into LLM reasoning without retraining.
A new arXiv paper introduces Magenta, a training-free agentic pipeline that converts natural-language math problems into Lean 4 statements and machine-checked proofs. It reports 100% accuracy on AIME 2025, AIME 2026, and HMMT February 2026, and solves all six IMO 2026 problems when paired with the K2-Horizon-7B reasoner. The work matters to practitioners because it shows formal verification signals can be integrated into LLM reasoning without retraining.
0 engagement1 sourcearxivThu, Sep 10, 2026, 09:48 AM
Social·Community
Benchmark shows LLaDA 8B diffusion LLM at 1.05 tokens/s on MacBook
A Medium article benchmarks LLaDA 8B, a diffusion-based LLM, against Llama 3 8B on a single MacBook with identical quantization and prompts. The author reports LLaDA achieves only 1.05 tokens per second, far below the claimed 1,000 tokens per second, and analyzes why the benchmark tool's own numbers may be misleading.
A Medium article benchmarks LLaDA 8B, a diffusion-based LLM, against Llama 3 8B on a single MacBook with identical quantization and prompts. The author reports LLaDA achieves only 1.05 tokens per second, far below the claimed 1,000 tokens per second, and analyzes why the benchmark tool's own numbers may be misleading.
0 engagement1 sourcerssSat, Aug 22, 2026, 06:03 PM
Releases·Model Release
IFM releases Jais 3 open model family, sizes 0.9B–375B, focused on Arabic and agentic tasks
IFM released Jais 3, an open-weight model family focused on Arabic and strong on agentic tasks. Sizes range from 0.9B to 375B, with GGUF quantizations of the smaller tiers already on Hugging Face. It matters to practitioners seeking open models with Arabic strength and local-run capability.
IFM released Jais 3, an open-weight model family focused on Arabic and strong on agentic tasks. Sizes range from 0.9B to 375B, with GGUF quantizations of the smaller tiers already on Hugging Face. It matters to practitioners seeking open models with Arabic strength and local-run capability.
73 engagement2 sourcesredditThu, Oct 15, 2026, 08:08 AM