Tencent Hy4 vs Mistral Large 4: Open-Weight Flagships a Week Apart, One's Weights Out and the Other's Pending, Output Prices $0.401 Apart

The self-hosting path is where the two models diverge most right now: Large 4's weight files are already downloadable, while the full Hy4 can currently only be called through hosted APIs, with no date set for its weight release. Hy4 preview weights have been public since Aug 28, but the preview and the full version are not the same checkpoint, and the full version's weights and complete system card are both marked as coming later.

Their positioning differs: the full Hy4 adds vision understanding and computer-use capabilities, with targeted training covering Blender, CAD and long multi-application workflows, plus game development and real financial tasks; Large 4 is aimed at coding and defensive cybersecurity, with Mistral presenting it as the leading open-weight coding model and publishing cybersecurity scores for vulnerability discovery, reproduction and patching.

Pricing moves in different directions: the full Hy4 keeps the preview's API pricing, at $0.834 input and $2.501 output per million tokens; Large 4's pricing is halved from the preview, with input falling from $1.36 to $0.70 and output from $4.18 to $2.10, and cached input at $0.07.

A note on comparability: the full Hy4 release does not restate its parameters, so the figures below carry over the preview's specifications and say so; Large 4's weight release adds about three weeks of reinforcement learning over the preview, and its coding scores moved accordingly, so preview scores and weight-release scores must not be mixed. Every figure below names its source and measurer.

1. Maker and release date: Hy4 is developed by Tencent's Hy team: the preview was released and open-sourced on Aug 28, 2026, and the full version was released on Nov 4, 2026; Large 4 is developed by Mistral: public preview on Oct 6, 2026, with improved weights released on Oct 29, 2026.
2. Total and active parameters: Hy4's full version does not restate its parameters; the preview's figures are 770 billion total and 49 billion active per token; Large 4's official blog says about 1 trillion parameters, and its model documentation gives 1.05 trillion total and 52 billion active per token (including embeddings and output layers; 49 billion excluding them).
3. Context window: Hy4's preview figure is over 1M tokens, and OpenRouter's deployment page lists 1,048,576 tokens with a 65,536-token maximum output; Large 4 is 1M tokens.
4. API input and output price: Hy4: $0.834 input and $2.501 output per million tokens; Large 4: $0.70 input and $2.10 output per million tokens; gaps of $0.134 on input and $0.401 on output.
5. Cached input price: Hy4: $0.042 per million cached input tokens; Large 4: $0.07 per million cached input tokens.
6. Weights and licence: Hy4 preview weights are public under the Apache 2.0 licence, while the full version's weights are planned for later release with no date given, as is its complete system card; Large 4 weights are public on Hugging Face (FP8 and NVFP4 versions) under the Mistral Ecosystem Open-Weight Licence, a modified MIT licence, under which a business offering inference as a service needs a separate commercial licence from Mistral.
7. Input modalities: Hy4 accepts text and image input and outputs text; the full version adds vision understanding and computer use, reading charts and documents, reasoning across multiple images, and operating command-line and graphical interfaces; Large 4 accepts text and image input and outputs text, is natively multimodal, and offers low, medium and high reasoning modes, with high as the default.
8. Access channels and subscription: Hy4 is hosted through Tencent Cloud TokenHub, OpenRouter and WorkBuddy, CodeBuddy, Yuanbao and ima, with a limited free allowance on WorkBuddy and CodeBuddy from Nov 4 to 15, and no standalone subscription price published; Large 4 is available through Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints, with weights available for self-hosted deployment, and no standalone subscription price published.
9. Regional availability: Hy4's OpenRouter deployment page lists Tencent Cloud as its only provider and notes that EU in-region routing is not currently available; Large 4's per-channel regional coverage has not been published item by item.

1. Independent index: Artificial Analysis Intelligence Index v5 (assessed Oct 29, 2026): Large 4 at high reasoning scores 26, at $2.06 per index task, down about 63% from the preview's $5.54; Hy4 has no published score on that index.
2. Coding: Published by Mistral: Large 4 scores 41% on Terminal-Bench 4.0, up 13 points from the preview; 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens per task; and 19.7% on Cognition's private FrontierCode 2.0 set. The full Hy4 has not published Terminal-Bench or SWE-bench scores.
3. Cybersecurity: Published by Mistral: Large 4 scores 51% on the Artificial Analysis Cyber Index, resists 97.1% of attacks on the B3 Agent Security Benchmark, and refuses 96.6% of malicious cyber requests on average. The full Hy4 limited its cybersecurity training data to defensive content, and its related benchmark scores have not been published.
4. Hy4 vendor-run evaluations: Published by Tencent: in a blind finance evaluation with 213 finance professionals and 579 real-world tasks, Hy4 scores 3.04, against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3; GameDev improves 19.1% over the preview and GameLogic 9.8%; in internal search evaluations the hallucination rate falls from 15.76% to 5.45% and speculative errors from 23.1% to 5.6%.
5. Large 4 other vendor-reported scores: Published by Mistral: BrowseComp 84.5% and Finch financial spreadsheet tasks 73.3%; on Artificial Analysis' AA-Omniscience v1.1 its hallucination rate is 28%, against 41% for the preview, with accuracy at 26%.
6. Not published, and comparison limits: There is no same-configuration score for Hy4 and Large 4 on a shared public benchmark; Hy4's finance blind evaluation and Large 4's Finch score come from different evaluations and are not directly comparable; each vendor's own benchmarks use its own methodology, and no further cross-model scores have been published.

1. For downloadable weights and self-hosting: If the task is deploying, customizing and fine-tuning a model on your own infrastructure, Large 4 fits, with its FP8 and NVFP4 weights public on Hugging Face.
2. For vision and computer use through a hosted API: If the task is reading charts and documents, reasoning across multiple images and operating command-line and graphical interfaces through a hosted API, the full Hy4 fits, with both capabilities available through TokenHub and OpenRouter since Nov 4.
3. For lower API list pricing: On list pricing, Large 4 fits, at $0.70 input and $2.10 output per million tokens, against Hy4's $0.834 and $2.501; for lower cached-input pricing, Hy4 fits, at $0.042 per million tokens against Large 4's $0.07.
4. For context above 1M tokens: If context needs exceed 1M tokens, Hy4 fits, with OpenRouter's deployment page listing 1,048,576 tokens; for a 1M-token context, Large 4 fits.
5. For published terminal-coding scores: If Terminal-Bench-style terminal tasks are the reference, Large 4 fits, with a published 41%; if a blind evaluation on real financial tasks is the reference, Hy4 fits, with a Tencent-run score of 3.04.
6. For already-public weights under Apache 2.0: If the task requires weights that are already public under the Apache 2.0 licence, Hy4 preview weights fit; the full version's weights are still pending and their licence has not been restated with the full release.