LIVE

Task-matched model directory · 12 current matches

Open-Source Alternatives to DeepSeek

Looking beyond DeepSeek? Browse open-weight candidates with overlapping coding, reasoning, and research tags. “Alternative” here means worth testing for a similar task—not equal benchmark performance.

These are potential alternatives based on published directory tags, not verified equivalents. Model quality depends on the task, prompt, runtime, quantization, and hardware. Public weights do not automatically mean an OSI-approved open-source license.

How to choose an alternative

DeepSeek includes different model families and sizes. Match the exact job, context length, hardware budget, and license before comparing; benchmark claims are only comparable when the test setup is disclosed.

This comparison is scoped to these directory categories: coding, reasoning, research, agents. Check the model card for the actual task benchmark, supported inputs, weight availability, license, and deployment requirements.

Check DeepSeek official site ↗

From synced Hugging Face and Ollama records

Models to evaluate

Search all models →

Hugging Face · directory match

Muse-Glimmer-30B-Abliterated-GGUF

Muse-Glimmer-30B-Abliterated-GGUF is a 16-bit quantized upload intended for image-text-to-text use, with both image and text modalities. The supplied Hugging Face metadata describes the model as having 28B parameters, although the model name and directory record use “30B”; this discrepancy should be verified before relying on the parameter count for hardware planning. Its listed capabilities and use cases are agentic workflows, agents, vision, and abliterated operation. The “Abliterated” classification is explicit in the model name, but the supplied material does not provide safety behavior details, evaluation results, or inference performance. No context-window value, tokens-per-second benchmark, pricing, or license is available in the supplied material. Because this is a GGUF quantized upload, its runtime behavior and memory requirements may differ from the underlying base model; the excerpt does not identify that base model or provide runtime configuration guidance. The supplied page content is for The-freezer/Muse-Glimmer-30B-Abliterated-GGUF, while the directory record points to Blackfrost-AI/Muse-Glimmer-30B-Abliterated-GGUF, so repository identity and provenance should also be confirmed before deployment.

agents
Size
30B parameters
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

Qwen3.8-Flash-Next-Uncensored-NVFP4

Qwen3.8-Flash-Next-Uncensored-NVFP4 is a 4-bit quantized Hugging Face upload presented as an image-text-to-text model with 120 billion parameters. The page identifies it as an image-and-text model, making it relevant for vision-oriented reasoning and multimodal text generation. The record also lists tool-calling, uncensored, and abliterated capabilities, so potential use cases include tool-calling workflows, vision tasks, and reasoning applications where those classifications are appropriate. This is a quantized upload rather than an indication that the underlying base model itself was originally trained in 4-bit precision; the supplied material does not identify the base model’s training or quantization process. The page does not provide a context-window value, inference speed benchmark, pricing, runtime requirements, or implementation guidance, so memory needs and expected throughput remain unverified. A license is not available in the supplied material for this record. Developers should also note that the source excerpt provides a concise model description but no task benchmarks, quality comparisons, or details about supported serving stacks. Tool-calling and multimodal support should therefore be validated in the actual files and runtime before production use.

reasoning
Size
Size not listed
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

Andy-4.2

Andy-4.2 is a 10-billion-parameter machine learning model hosted on Hugging Face. The available record positions it for agents, agentic workflows, and tool-calling, making it potentially relevant for applications that need a model to interact with tools or participate in multi-step workflows. However, the supplied material does not identify its model task, architecture, supported input or output modalities, context window, quantization format, inference requirements, or measured tokens-per-second performance. It also does not provide benchmarks, examples, or implementation guidance that would help evaluate quality for a specific workload. No license is available in the supplied material for this record, so redistribution and commercial-use terms remain unverified. The Hugging Face page describes the repository as a downloadable model that can be explored and run on Hugging Face, but it does not establish a hosted API, pricing, or runtime defaults. Developers should therefore treat the agentic and tool-calling classifications as the main advertised use-case signals and verify compatibility, licensing, and deployment details from the model repository before adoption.

agents
Size
Size not listed
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

Qwen3.8-27B-Turbo-Brilliance-50X-Reasoning-Instruct-MTP-MLX-8bit

Qwen3.8-27B-Turbo-Brilliance-50X-Reasoning-Instruct-MTP-MLX-8bit is a Hugging Face image-text-to-text repository intended for multimodal vision and reasoning workloads. The record lists 27,356,728,560 parameters and text and image modalities, making it substantially larger than typical lightweight local models and potentially relevant for applications that need image understanding alongside language generation. Its stated use cases are vision and reasoning; no additional capabilities, such as tool calling, coding, or agent support, are identified in the supplied material. The repository name identifies an MLX-8bit upload, distinguishing it from an unquantized base-model release, but the structured quantization field is not populated, so the exact quantization implementation and runtime requirements are not verified here. The model task is explicitly image-text-to-text, rather than image generation or text-only generation. An advertised context window, inference speed benchmark, pricing, and deployment requirements are not available in the supplied material. This specific record is listed under the Apache-2.0 license; confirm that license terms and any applicable restrictions fit your intended deployment before use.

reasoning
Size
27.4B parameters
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

qwen3.8-27b-aeon-ultimate-uncensored-attention8-bf16recurrence-vision-mtplx

qwen3.8-27b-aeon-ultimate-uncensored-attention8-bf16recurrence-vision-mtplx is a 4-bit quantized Hugging Face upload presented as an image-text-to-text model with 27 billion parameters. Its listed modalities are image and text, making it relevant for multimodal applications such as image understanding, visual question answering, and text generation, although the supplied material does not provide task-specific examples or evaluation results. Tool-calling and uncensored behavior are listed capabilities, and reasoning is listed as a use case; these classifications should be treated as advertised metadata rather than verified performance claims. The 4-bit format distinguishes this upload from any higher-precision base model and may reduce hardware requirements, but the excerpt does not specify memory usage, supported runtimes, hardware recommendations, or inference speed. An advertised context window is not available in the supplied material, and no tokens-per-second benchmark is provided. The license for this specific record is also not available in the supplied material, so redistribution and commercial-use terms require confirmation from the model card or repository files before deployment.

reasoning
Size
27B parameters
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

reasoning-quantization

reasoning-quantization is a Hugging Face model repository presented for reasoning-related use, but the supplied material does not provide enough technical detail to validate it for deployment. The repository page identifies it as a machine learning model and offers download, file exploration, and Hugging Face execution, while the model record lists reasoning as its use case. No model task, capabilities, or modalities are specified, so supported input and output types, tool use, coding ability, and agent workflows remain unverified. Although the repository name includes “quantization,” the record does not identify the quantization method, numeric format, supported hardware, runtime requirements, or the base model from which this upload was derived. The context window, parameter count, inference speed, pricing, and benchmark results are also unavailable in the supplied material. A license is not specified for this record, so redistribution and commercial-use terms should be confirmed before adoption. Developers should therefore treat this entry as a minimally documented repository rather than a fully characterized model release and inspect its files and documentation directly before integrating it.

reasoning
Size
Size not listed
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

Qwen3.8-Flash-Next-Abliterated-Uncensored-Turbo-Brilliance-50x-Reasoning-Instruct-GSQ-RCO-GGUF

Qwen3.8-Flash-Next-Abliterated-Uncensored-Turbo-Brilliance-50x-Reasoning-Instruct-GSQ-RCO-GGUF is a Hugging Face image-text-to-text model repository intended for multimodal text and image inputs. The supplied Hugging Face metadata identifies it as a 177-billion-parameter model, making it a substantially sized model for developers evaluating vision and reasoning workloads. Its name and record explicitly classify it as abliterated and uncensored, while the listed use cases also include vision and reasoning. The repository name ends in GGUF, but the supplied record does not provide a specific quantization level or other quantization details, so the exact storage or runtime format should be verified before deployment. No context-window size, pricing, or inference-speed benchmark is available in the supplied material. This record is licensed under Apache-2.0. Developers should therefore treat it as a multimodal model suitable for image-and-text understanding, reasoning-oriented prompting, and applications requiring the repository’s stated uncensored or abliterated characteristics. The available metadata does not establish performance, hardware requirements, serving compatibility, safety behavior beyond those classifications, or whether a particular GGUF file is appropriate for a target runtime.

reasoning
Size
Size not listed
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

Qwen3.8-27B-Turbo-Brilliance-50X-Reasoning-Instruct-NEO-MAX-MTP-GGUF

Qwen3.8-27B-Turbo-Brilliance-50X-Reasoning-Instruct-NEO-MAX-MTP-GGUF is a 27-billion-parameter Hugging Face repository intended for image-text-to-text workloads. It accepts text and image inputs, making it potentially useful for vision tasks that also require instruction following or reasoning, such as interpreting visual content alongside written prompts. The repository metadata identifies its task as image-text-to-text and lists vision and reasoning as practical use cases. The name includes “GGUF,” indicating an upload associated with the GGUF format, but the supplied material does not verify a specific quantization level or provide runtime configuration details. This record is licensed under Apache-2.0. Developers should treat the model’s context window, inference speed, pricing, and detailed capabilities as unverified because those values are not available in the supplied material. No benchmark results, tool-calling support, coding specialization, or deployment requirements are provided. The source page describes the repository as a place to download, explore files, and run the model on Hugging Face, but the excerpt does not establish compatibility with a particular local inference engine or hardware profile. It should therefore be evaluated for the intended multimodal reasoning workflow before production use.

reasoning
Size
27B parameters
License tag
Check source card
Verify model card ↗

Ollama · directory match

dolphin3:4.1M

dolphin3:4.1M is listed as a Dolphin 3.0 Llama 3.1 8B instruct-tuned model intended as a general-purpose local model for coding, mathematics, agentic workflows, function calling, and general use. The supplied record also associates it with coding, agents, agentic tasks, tool-calling, and reasoning. Its listed parameter count is 4,100,000, although that figure is inconsistent with the “Llama 3.1 8B” description and should be verified before making hardware or performance decisions. The supplied source URL currently returns an Ollama 404 page, so the model card and its technical claims cannot be independently confirmed from the provided material. No model task classification, license, quantization information, context window, modalities, capabilities list, pricing, or inference-speed benchmark is available in the supplied material. In particular, there is no verified evidence here for a specific context length, runtime configuration, tool schema, or supported input/output modality. The record characterizes the model as suitable for local coding and reasoning applications, but provides no benchmark data to assess quality, latency, memory requirements, or comparative performance. Licensing should be treated as unresolved for this exact record until a valid model page or authoritative documentation is available.

codingagentsreasoning
Size
4M parameters
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

Qwen3.8-27B-Brainwaves-2M-qx64-hi-mlx

Qwen3.8-27B-Brainwaves-2M-qx64-hi-mlx is a 6-bit quantized Hugging Face upload presented as an image-text-to-text model with 27 billion parameters. Its listed modalities are image and text, making it relevant to applications that combine visual inputs with language generation. The supplied use cases are coding, vision, reasoning, and research, so it may be worth evaluating for multimodal developer tools, visual analysis, technical question answering, and exploratory research workflows. This record describes the quantized upload, not the specifications or behavior of an underlying base model; compatibility, tokenizer details, and recommended inference software are not available in the supplied material. No context-window value, inference-speed benchmark, pricing, or additional capability metadata is provided, so latency, maximum prompt size, and production suitability remain unverified. The source excerpt also does not provide a license for this specific record; review the model card and repository files before redistribution or commercial deployment. Because the available description is brief, safety behavior, image-resolution support, tool-calling support, and quality relative to other 27B models should be tested rather than assumed.

codingreasoningresearch
Size
27B parameters
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

nanok-midtrain-agentic-v2-100B

nanok-midtrain-agentic-v2-100B is presented on Hugging Face as a machine-learning model intended for agentic workflows, but the supplied material leaves several important implementation details unresolved. The model record identifies agents and agentic use cases and lists agentic capability; no modalities are specified, so the available material does not verify text, image, audio, or other input/output support. The Hugging Face page metadata describes the exact repository as having 32B parameters, while the model record labels it as 100B based on the model name, creating a material parameter-count discrepancy. Consequently, the usable parameter count should be verified before deployment. Model task, context window, quantization, license, inference speed, pricing, and benchmark results are not available in the supplied material. There is also no evidence here of tool-calling support beyond the general agentic classification, nor of a published model card explaining training data, hardware requirements, safety behavior, or recommended runtime settings. This record is therefore most useful as a pointer to an agent-oriented Hugging Face repository, not as a complete basis for production selection.

agents
Size
100B parameters
License tag
Check source card
Verify model card ↗

Hugging Face · directory match

gemma4-comfyui-mcp

gemma4-comfyui-mcp is a Hugging Face text-generation model with 12 billion parameters, intended for agent and tool-calling workflows. The record identifies text as its modality and lists tool-calling as its capability, making it potentially relevant for applications that need a language model to select or invoke external tools. However, the supplied material does not explain the specific ComfyUI or MCP integration, tool schema, prompting format, supported runtimes, or whether additional configuration is required. The model’s context window, inference speed, pricing, and benchmark results are not available in the supplied material. Quantization is also not specified, so this record cannot establish whether the repository contains a quantized upload or an unquantized/base model. A license is not available in the supplied material for this record and should be verified before redistribution or commercial deployment. The source page describes the repository as a Text Generation model with 12B parameters; no separate evidence is provided for vision, audio, image generation, coding specialization, or other modalities. Developers should therefore treat its practical scope as text-based generation with advertised tool-calling support, pending verification of the repository files and model-card instructions.

agents
Size
Size not listed
License tag
Check source card
Verify model card ↗

Related tools