AI models for code, context, media, and local use

Ornith spans checkpoints from 9B to 397B, with GGUF and FP8 variants, so the coding-model choice can account for both scale and a realistic serving format.

North Mini Code activates 3B of its 30B parameters and targets terminal work, putting the active-compute tradeoff directly into an agentic coding decision.

Mellum2 pairs a software-engineering focus with 131K context notes and several checkpoints, a useful combination when repository length may rule out a smaller coding model.

GLM-5.2 brings 1M-context positioning and local-serving paths to long-horizon coding, putting deployment alongside the question of whether the project can stay in context.

Kimi-K2.7-Code keeps reasoning across tool calls and offers an INT4 route, exposing the choice between multi-step execution needs and a lower-bit serving option.

Kimi-K2.6 covers visual input as well as long-horizon coding and tool use, adding a comparison point for work that moves between screenshots, code, and commands.

Qwen3.6-35B-A3B combines open weights, multimodal input, tool use, and long context, bringing local weight access into the software-workflow decision.

Qwen-AgentWorld supplies simulated MCP, terminal, web, Android, and OS environments, shifting the question from code completion to behaviour inside interactive software settings.

LongCat 2.0 combines coding, long-context, and agent-oriented work with base, FP8, and INT8 model pages, keeping both the software task and serving format in the comparison.

MiniMax-M3 puts 1M context, multimodal input, and separate thinking modes in one model, clarifying the tradeoff between a very large workspace and controllable reasoning behaviour.

DeepSeek-V4 brings long-context reasoning, coding, and agentic evaluation into one family, giving several multi-step workloads a shared-model comparison point.

At 295B parameters with public weights and serving notes, Hy3 makes the infrastructure cost of a large deployable reasoning model part of the choice.

Step-3.7-Flash pairs 256K context with BF16, FP8, NVFP4, and GGUF variants, letting context requirements be weighed against different serving footprints.

Command A+ W4A4 applies low-bit quantization to a long-context, multimodal tool-use model, bringing hardware requirements into a broader agent-workload decision.

Trinity-Large-Thinking centres coherent multi-turn behaviour, constrained instructions, and clean tool use for long interactions that need the model to follow the same task and rules across multiple turns.

Ling-2.6-1T makes the scale decision explicit: demanding multi-step reasoning comes with trillion-parameter weights, multi-GPU serving notes, or a hosted access route.

Gemma 4’s 12B dense checkpoint handles visual and audio input, covering local fallback jobs that need more than text without a separate encoder stack.

At 230M parameters with GGUF, ONNX, and MLX variants, LFM2.5-230M offers a compact route for testing extraction or lightweight agent work on-device.

The 350M checkpoint provides a direct step beyond the family’s 230M option, helping readers judge how much on-device model capacity their task needs before moving to billion-parameter hardware.

LFM2.5-8B-A1B activates 1.5B of its 8.3B parameters, putting the balance between tool-use capacity and active on-device compute in view.

MiniCPM5-1B combines a 1B-class footprint with 131K context and several local runtimes, exposing the tradeoff between a compact assistant and long inputs.

ZAYA1-8B activates 760M of 8.4B parameters, making reasoning capacity versus per-token activation the central local-model tradeoff.

Inkling accepts text, images, and audio with adjustable reasoning effort, making cross-media input and reasoning control the comparison.

Nemotron 3.5 Content Safety separates moderation into its own model layer, with prompt, image, and response checks against standard or custom policies.

MOSS-VL separates continuous video, offline multimodal work, and further training across Realtime, Instruct, and Base checkpoints, making the intended video workflow easier to choose.

TIPS is an encoder rather than a full conversational model, providing reusable image-text and spatial representations as an upstream layer.

DeepSeek-OCR-2 targets image and PDF conversion to Markdown with dynamic resolution, putting document extraction ahead of general visual conversation.

GLM-OCR focuses the choice on complex document parsing and deployment efficiency, rather than asking a general multimodal model to handle structured pages.

Unlimited-OCR handles long documents in one shot and supports batch image and PDF inference, addressing workloads that may strain page-by-page OCR.

Lance covers image and video understanding, generation, and editing in one workflow; its stated 40GB VRAM requirement keeps the hardware tradeoff visible.

TimesFM brings forecast-specific checkpoints and a supported BigQuery ML path to time-series work, rather than relying on general language reasoning.

Isaac GR00T N1.7 connects perception and instructions to robot actions through a vision-language-action model and LeRobot-format fine-tuning workflow.

OLMoEarth is trained around satellite imagery, band extraction, and planetary-scale mapping, matching Earth-observation work that needs domain-specific representations.

Hy-MT1.5 combines 33-language offline translation with a 1.25-bit GGUF route, bringing multilingual coverage into an Android or similarly constrained device decision.

LFM JP groups Japanese-tuned text, tool-use, structured-output, transcription, speech-generation, and speech-to-speech models for workflows where Japanese language coverage is the specialist requirement.