Open Models

Training-data curation can improve model quality and inference efficiency without simply adding compute. The practical work is decontamination, deduplication, balancing, task matching, and selective synthesis.

MiniMax and Together describe why agent workloads change inference priorities: large codebase context, repeated tool turns, multimodality, and long-horizon RL all demand purpose-built environments and kernels.

β-OPSD exposes self-distillation’s fixed regularization as a tunable parameter, then approximates policy optimization through logit mixing. It targets more stable reasoning training without direct RL.

Hugging Face’s Hub shows how to scale model search: precompute tokens, separate metadata from artifacts, isolate heavy reads, and autoscale from workload signals.

DataOrchestra chooses a processing pipeline per pre-training example, improving average benchmark results while avoiding compute on chunks that need no transformation.

poolside treats synthetic data and training observability as one scaling problem: diversify scarce code data, decompose hard generation tasks, and verify numerical invariants before failures compound.

Edge-agent design is often memory-bound: use small prompted models when devices can spare gigabytes, or fine-tune tiny models for narrow functions when reach and latency matter more.

This security eval tests whether agents can discover and exploit logic flaws across live chained services, using hidden zero-days and deterministic grading instead of source-code pattern matching.

VLM-IE3D adds implicit and reconstructed geometry tokens to an RGB-video VLM, offering an open approach for agents that must reason about spatial scenes without dedicated 3D input.

Agent economics can regress even when token prices look stable. Route by task, preserve model optionality, and move deterministic work out of LLM calls before scaling usage.

Ling 3.0 Flash joins AI Gateway with a 256K context window, thinking and non-thinking modes, and free access through August 3 for agent workload testing.

Poolside’s open-weight Laguna S 2.1 is now callable through Vercel AI Gateway in free 256K and paid 1M-context variants, with thinking and no-thinking modes for coding agents.

Hugging Face demos real-time speech-to-speech from open parts: Nvidia Parakeet ASR, Gemma 4 31B on Cerebras inference, Alibaba's Qwen3TTS — pipeline code is open and already runs on 9,000+ Reachy Mini robots.

Program-as-Weights compiles natural-language fuzzy functions (JSON repair, log filtering) into adapters for a frozen 0.6B interpreter — matching Qwen3-32B prompting at ~1/50th the memory, 30 tok/s on an M3.

Local-first meeting assistant: Whisper/Parakeet transcription with a 4x speed claim and Ollama summaries, all on-device via a Rust/Tauri app — a working template for shipping private, local AI features.

CompactionRL uses RL to teach agents to compact their own context mid-task, lifting GLM-4.5-Air 7 points to 66.8% on SWE-bench Verified; the recipe is now in GLM-5.2's training pipeline.

Google's June roundup: Gemma 4 12B runs locally in 16GB of memory, Gemini 3.5 Flash adds computer use for desktop, mobile, and browser agents, and Nano Banana 2 Lite ships as a cheaper image model.

A free, code-oriented Chinese curriculum spans model tuning, deployment, agents, alignment, security, and multimodal systems. It is useful as a broad learning map, but remains a work in progress.

Vercel’s June gateway data shows cheap volume moving to open-weight models while costly agent workloads stay on frontier models, strengthening the case for risk-based routing.

NVIDIA's NeMo AutoModel makes MoE fine-tuning ~3.4-3.7x faster on Transformers v5 while keeping the AutoModelForCausalLM API — relevant if you tune open models like Qwen3-30B-A3B for your own agents.

HF Jobs now stands up an OpenAI-compatible vLLM endpoint in one command, billed per second (A10G at $1.50/hr). Useful for throwaway endpoints: one-off evals, batch runs, agent experiments against open models.