Qwen3.5

2026-10-25

MODELS · 1 SOURCE · The Decoder

Nvidia open-sources Flora, a Qwen3.5-based robot video labeling engine

Nvidia's GEAR lab open-sourced Flora, a data-engine composed of three fine-tuned Qwen3.5 models (9B for segmentation, 27B for descriptions, and a 27B 'Flora-Rewarder' for scoring) that labels robot videos for action boundaries, step-level rewards, and task success. Nvidia says Flora is already used in training its GR00T models at scale on GB300 GPUs and estimates labeling costs of about $0.72/hour versus ~$88/hour with GPT-6 Astra, though the cost figures are projections and benchmark comparisons include methodological caveats.

7.0

2026-09-25

MODELS · 1 SOURCE · arXiv cs.AI

Pistis: 27B and 9B multimodal LLM family with IDRL post-training (arXiv:2609.28554v1)

arXiv:2609.28554v1 presents the Pistis model family—27B- and 9B-parameter multimodal LLMs built on Qwen3.6 and Qwen3.5—trained via a scalable post-training framework that starts with large-scale multimodal supervised fine-tuning and then applies Interleaved Distillation and Reinforcement Learning (IDRL). The paper describes two specialized variants (Pistis-Thinking for deep multimodal reasoning and Pistis-Agentic for long-horizon planning and tool use), plus a system method Pistis-Auto-Harnessing (PAH); both scales reportedly outperform their corresponding base models, with the Agentic variant strong in multimodal search.

7.0

2026-09-23

CODING · 1 SOURCE · NVIDIA Developer

SWE-Serve benchmark shows many patches that pass local tests fail live serving for SGLang

SWE-Serve is a new benchmark that turns 83 merged SGLang pull requests into 53 executable inference-serving tasks across six families; 19 tasks start a real server and validate patches through the full serving path. Across those live-serving tasks, the same patches pass 45.9% of the time with the complete verifier versus 69.4% when live-serving checks are excluded (roughly one in three otherwise-passing patches fail live checks); multi-domain tasks show a 21.3 percentage-point lower pass rate, and closed-book model pass@1 scores (11 models, mini-swe-agent) ranged from 34.6%–75.5% with Claude Opus 5 and GPT-5.6 Sol among top performers.

7.0

2026-09-22

MODELS · 1 SOURCE · Hugging Face

Transformers adds GGUF (llama.cpp quant) support on main

Hugging Face's transformers (main branch) can now load and run GGUF quantized checkpoints using llama.cpp/ggml kernels, aiming for efficient local inference (initially optimized for Apple Silicon and Qwen3.5 models). The integration also enables serving GGUF checkpoints via transformers-serve and falls back to dequantization if compatible kernels are unavailable.

7.0

2026-09-16

MODELS · 1 SOURCE · Cohere

Cohere argues smaller, task‑optimized language models can outperform at lower cost

Cohere published a post promoting small language models (SLMs) for enterprise use, arguing they cut compute, data and energy needs while enabling local deployment and lower costs. The post highlights Cohere models — Command R7B, the Tiny Aya family (3.35B) and North Mini Code (30B MoE with 3B active) — and cites benchmark claims that North Mini Code and Tiny Aya outperform several larger competitors on coding and multilingual translation benchmarks respectively.

6.0

2026-05-22

MODELS · 1 SOURCE · Mistral AI

Mistral launches Medium 3.5 (128B) open-weights in public preview and adds cloud remote agents to Vibe/Le Chat

Mistral Medium 3.5, a 128B dense ‘merged’ model with a 256k context window and open weights under a modified MIT license, is available in public preview and becomes the default in Le Chat and Vibe coding flows. The update also introduces cloud remote agents for Mistral Vibe and a new Work mode in Le Chat for multi-step tool-driven tasks; pricing is $1.5 per million input tokens and $7.5 per million output tokens, and the model claims strong benchmark and agentic scores (e.g., 77.6% SWE-Bench Verified).

8.0