Skip to content
TheChatArchiveThe Chat Archive
Perplexitysonar· PerplexityHugging Face LMArena Search Arena 24K

1. llama-4-maverick-17b-128e-instruct 2. gemma-3-12b-it 3. gemma-3-4b-it 4.…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Perplexitysonar-pro-high’s answer.
  1. User
    1. llama-4-maverick-17b-128e-instruct
    2. gemma-3-12b-it
    3. gemma-3-4b-it
    4. doubao-1.5-pro-32k-250115
    5. doubao-1.5-vision-pro-32k-250115
    6. cobalt-exp-beta-v3
    7. cobalt-exp-beta-v4
  2. Perplexity Assistant

    Here is a detailed overview of the listed models based on the latest available information:

    1. Llama-4-Maverick-17B-128E-Instruct

    • Developed by Meta as part of the Llama 4 series.
    • Parameters: 17 billion active parameters with 128 experts (total 400 billion parameters).
    • Multimodal capabilities: supports multilingual text, image, and code understanding.
    • Context window: very large, supporting up to 128k tokens.
    • Performance: Outperforms GPT-4o and Gemini 2.0 Flash on several benchmarks, with strong scores on GPQA Diamond (69.8) and MMLU Pro (74.3).
    • Image grounding: best-in-class ability to align user prompts with visual concepts and anchor responses to image regions.
    • Released April 2025, available with BF16 and FP8 quantized weights for efficient deployment.
    • Usage: Hugging Face integration with transformers library support.
      [3][6]

    2. Gemma-3-12B-IT

    • Part of Google's Gemma 3 family, a state-of-the-art open model series.
    • Parameters: 12 billion.
    • Instruction-tuned (IT) version optimized for chat and instruction following.
    • Multimodal: supports both text and images.
    • Context window: up to 128k tokens.
    • Multilingual: supports over 140 languages.
    • Enhanced reasoning, math, coding, and chat capabilities.
    • Uses a new tokenizer and vision encoder based on SigLIP.
    • Outperforms previous Gemma 2 models and competes well with Gemini 1.5-Pro.
      [1][4][7]

    3. Gemma-3-4B-IT

    • Smaller variant of Gemma 3 instruction-tuned model with 4 billion parameters.
    • Also multimodal with text and image input support.
    • Context window: 128k tokens.
    • Supports 140+ languages.
    • Offers improved math, reasoning, and chat capabilities compared to earlier versions.
    • Available freely on platforms like OpenRouter and Hugging Face.
      [1][4][7]

    4. Doubao-1.5-pro-32k-250115

    • Developed by ByteDance, released early 2025.
    • Uses a sparse Mixture-of-Experts (MoE) architecture for efficiency.
    • Activated parameters equivalent to 20B dense model but with much lower computational cost.
    • Context window: 32k tokens (extended up to 256k tokens in some configurations).
    • Multimodal: supports vision, voice, and text.
    • Known for “Deep Thinking” mode enhancing reasoning capabilities.
    • Matches or surpasses GPT-4o and Claude 3.5 Sonnet on benchmarks like AIME.
    • Cost-effective with pricing significantly lower than competitors.
    • Strong presence in China with nearly 60 million monthly active users.
      [2][5][8]

    5. Doubao-1.5-vision-pro-32k-250115

    • Variant of Doubao-1.5-pro with enhanced vision capabilities.
    • Supports high-resolution image inputs and complex visual-textual reasoning.
    • Maintains the same efficient MoE architecture and extended context window.
    • Designed for applications requiring detailed image understanding and multimodal interaction.
      [2][5]

    6. Cobalt-exp-beta-v3

    • No direct detailed public information found in the current search results.
    • Likely an experimental or beta version of a model named Cobalt, possibly related to multimodal or instruction-tuned AI models.
    • Further details would require access to specific release notes or developer documentation not currently available.

    7. Cobalt-exp-beta-v4

    • Similar to v3, no detailed public information available from current sources.
    • Presumably a newer experimental beta iteration of the Cobalt model series.
    • Expected to include improvements or new features over v3, but specifics are not documented in the provided data.

    Summary Table

    ModelParamsMultimodalContext WindowLanguages SupportedNotable FeaturesRelease / Source
    Llama-4-Maverick-17B-128E-Instruct17B active (400B total)Yes128kMultilingual (200+ pretraining)Strong image grounding, high benchmark scoresMeta, April 2025 [3][6]
    Gemma-3-12B-IT12BYes128k140+Multimodal, instruction tuned, improved reasoningGoogle, March 2025 [1][4]
    Gemma-3-4B-IT4BYes128k140+Smaller, multimodal, instruction tunedGoogle, March 2025 [1][7]
    Doubao-1.5-pro-32k-250115~20B activated (MoE)Yes32k (up to 256k)MultilingualSparse MoE, deep thinking mode, cost-effectiveByteDance, Jan 2025 [2][5][8]
    Doubao-1.5-vision-pro-32k-250115Same as aboveYes32k (up to 256k)MultilingualEnhanced vision capabilitiesByteDance, Jan 2025 [2][5]
    Cobalt-exp-beta-v3UnknownUnknownUnknownUnknownExperimental beta versionNot publicly detailed
    Cobalt-exp-beta-v4UnknownUnknownUnknownUnknownExperimental beta version, newer iterationNot publicly detailed

    This overview captures the key characteristics and recent advancements of these models, highlighting their architectures, capabilities, and intended use cases. The Llama 4 Maverick and Gemma 3 series stand out for their large context windows and multimodal support, while Doubao 1.5 Pro emphasizes efficiency and cost-effectiveness with MoE technology. The Cobalt experimental models lack publicly available detailed information as of now.

  3. User

    Please put the models into a table with the following columns: Description Year of Release Exact Date Date notes Open Source? License Owner User Feedback (Caveats Apply)

  4. Perplexity Assistant

    Here is a table summarizing the requested details for the listed models based on the available information:

    ModelDescriptionYear of ReleaseExact DateDate NotesOpen Source?LicenseOwnerUser Feedback (Caveats Apply)
    llama-4-maverick-17b-128e-instructHigh-capacity multimodal MoE model with 17B active params, 128 experts, 1M token context window, instruction-tuned, strong vision-language capabilities20252025-04-05Released under Llama 4 Community LicenseYes (restricted)Llama 4 Community LicenseMeta AIHighly efficient, strong image grounding, requires large GPU resources; open source but with usage restrictions[5][8]
    gemma-3-12b-itGoogle’s 12B parameter instruction-tuned multimodal model supporting 140+ languages, 128k token context window2025Early 2025Announced Q1 2025NoProprietaryGoogle DeepMindStrong reasoning and multilingual support; not open source; advanced multimodal capabilities[1][4]
    gemma-3-4b-itSmaller 4B parameter variant of Gemma 3 instruction-tuned multimodal model2025Early 2025Announced Q1 2025NoProprietaryGoogle DeepMindSimilar capabilities to 12B version but smaller scale; not open source[1][4]
    doubao-1.5-pro-32k-250115ByteDance sparse MoE model with ~20B activated params, 32k token context, multimodal (vision, voice, text), cost-effective20252025-01-15Released January 2025NoProprietaryByteDanceEfficient, strong reasoning ("Deep Thinking" mode), large user base in China; proprietary[2][5][8]
    doubao-1.5-vision-pro-32k-250115Vision-enhanced variant of Doubao 1.5 Pro with extended multimodal capabilities and 32k token context window20252025-01-15Released January 2025NoProprietaryByteDanceEnhanced image understanding; same efficiency and cost benefits as Doubao 1.5 Pro; proprietary[2][5]
    cobalt-exp-beta-v3Experimental beta version of Cobalt model; details scarce, likely proprietaryUnknownUnknownNo public release date availableNoUnknownCobalt (unknown)Limited information; likely proprietary and experimental; no public user feedback available[3][6]
    cobalt-exp-beta-v4Newer experimental beta iteration of Cobalt model; details scarceUnknownUnknownNo public release date availableNoUnknownCobalt (unknown)Limited information; likely proprietary and experimental; no public user feedback available[3][6]

    Notes:

    • Open Source?: Llama 4 Maverick is released under a community license allowing research and commercial use with some restrictions; not fully open source in the traditional sense but accessible. Gemma 3 and Doubao models are proprietary. Cobalt models have no public open-source status.
    • License: Llama 4 Maverick uses the Llama 4 Community License; others are proprietary with no public license details.
    • User Feedback: Llama 4 Maverick praised for multimodal and large context capabilities but requires significant hardware. Gemma 3 models are strong in reasoning and multilingual tasks but closed source. Doubao models are noted for efficiency and cost-effectiveness in China. Cobalt models lack public feedback or detailed info.
    • Exact Dates: Some models have precise release dates (Llama 4 Maverick: April 5, 2025; Doubao: January 15, 2025). Gemma 3 models are early 2025 releases without exact day. Cobalt models have no public release dates.

    This table reflects the best available data as of April 2025.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.