MiniCPM-V
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 4,096 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 21, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 4,096 words
Add missing citations, update stale details, or suggest a clearer explanation.
MiniCPM-V is a family of open-weights multimodal large language models developed by the OpenBMB lab at Tsinghua University's Natural Language Processing group together with the spin-off company ModelBest Inc. The series targets efficient on-device deployment, packaging vision, OCR, video, and (since the MiniCPM-o branch) speech understanding into models small enough to run on a contemporary smartphone or tablet. The flagship August 2024 release, MiniCPM-Llama3-V 2.5, was billed by its authors as "the first end-side MLLM achieving GPT-4V level performance," with results above GPT-4V-1106, Gemini Pro, and Claude 3 on the OpenCompass aggregate of eleven vision benchmarks.[^1] The lineage stretches from a 3B-parameter pilot in January 2024 through MiniCPM-o 2.6, an 8B end-to-end omnimodal model released in January 2025 that adds real-time bilingual speech conversation.[^2][^3]
| Field | Value |
|---|---|
| Developer | OpenBMB / ModelBest Inc. / Tsinghua University NLP Lab[^1][^4] |
| First release | MiniCPM-V 1.0, February 2024[^5] |
| Latest covered release | MiniCPM-o 2.6, January 24, 2025[^3] |
| Parameter sizes | 2.8B, 3B, 8B (depending on variant)[^1][^5][^6] |
| Vision encoder | SigLIP SoViT-400m/14[^1] |
| Adaptive encoding | LLaVA-UHD style image slicing, perceiver resampler[^1][^7] |
| Max image resolution | 1.8M pixels at any aspect ratio[^1] |
| Key paper | arXiv:2408.01800 (Aug 3, 2024)[^1] |
| Code repository | github.com/OpenBMB/MiniCPM-V[^4] |
| License | Apache-2.0 (code); free commercial use after registration (weights)[^6] |
OpenBMB ("Open Lab for Big Model Base") is a research lab jointly operated by the Tsinghua University NLP group and ModelBest Inc., a spin-off founded in 2022 in Beijing's Haidian district by Tsinghua researchers.[^8] ModelBest positions itself around a thesis that small, distilled language and multimodal models can match much larger systems on practical tasks while remaining cheap enough to deploy on personal devices.[^8] In December 2024, MIT Technology Review reported that the company had closed a third funding round in the "tens of millions of dollars," part of a broader wave of Chinese on-device AI startups that emerged alongside DeepSeek.[^8]
The MiniCPM ("Mini Chinese-English Pre-trained Model") project began as the text-only MiniCPM-2B, released in February 2024, which the authors reported as performing comparably to Mistral-7B on public benchmarks despite using roughly a third of the parameters.[^9] A 4B-parameter MiniCPM3 followed in September 2024, with the authors claiming results above Phi-3.5-mini-instruct and GPT-3.5-Turbo-0125 and competitive with Qwen2-7B and 8B-class Llama 3 variants.[^9] The text-only line later iterated into MiniCPM4 (June 2025) and MiniCPM4.1 (September 2025), both emphasizing inference acceleration via trainable sparse attention (InfLLM-V2) and BitCPM-style quantization, but the vision branch, MiniCPM-V, predates and parallels that trajectory rather than depending on it directly.[^9] The vision lineage reused the same base language model in its first two releases and inherited the same on-device-first design philosophy.
The first MiniCPM-V model, sometimes referred to as OmniLMM-3B in early documentation, paired a SigLIP-400M vision encoder with the 2.4B-parameter MiniCPM text base through a perceiver-style resampler.[^5] The release emphasized aggressive token compression: image features were squeezed into 64 visual tokens, versus more than 512 tokens for typical MLP-projector models such as LLaVA.[^5] On general multimodal benchmarks the authors reported 1452 on MME, 67.9 on MMBench (English) and 37.2 on MMMU, ahead of the 9.6B-parameter Qwen-VL-Chat and the 17.4B-parameter CogVLM at similar settings.[^5] Although 1.0 lacked the adaptive resolution scheme that would come in 2.0, it already exhibited the family's signature design choice: privilege end-side deployability over raw parameter count, and rely on the resampler to keep visual token budgets small enough for mobile inference.[^5]
MiniCPM-V 2.0, released on April 12, 2024, kept the 2.8B-parameter footprint but introduced two changes that defined the rest of the family.[^6] First, it integrated the adaptive visual encoding scheme from LLaVA-UHD, allowing the model to accept images up to roughly 1.8 million pixels (for example 1344x1344) at any aspect ratio rather than forcing a fixed 336x336 square.[^6][^7] Second, it was the first end-side multimodal model from the group aligned with multimodal RLHF, drawing on the RLHF-V technique from the same authors.[^6][^10] The authors reported that 2.0 reached scene-text understanding comparable to Gemini Pro and surpassed Qwen-VL-Chat 9.6B, CogVLM-Chat 17.4B, and Yi-VL 34B on OCRBench, TextVQA, MME, MMBench, and MathVista.[^6] Importantly, 2.0 also matched the much larger GPT-4V on the Object HalBench hallucination benchmark, which the authors attributed primarily to the RLHF-V alignment step rather than to the visual encoder change.[^6] The model shipped on Hugging Face under an Apache-2.0 code license, with weights free for academic use and free for commercial use after registration, a licensing template carried through all subsequent releases.[^6]
Released on May 20, 2024, MiniCPM-Llama3-V 2.5 swapped the small MiniCPM base for Llama3-8B-Instruct, bringing the total parameter count to roughly 8B (8B LLM plus the 400M SigLIP encoder and a thin resampler).[^11] This was the version the authors positioned as the first open end-side model reaching GPT-4V-class quality: the OpenCompass average across eleven benchmarks rose to 65.1, ahead of GPT-4V-1106 at 63.5, with OCRBench above 700 (versus 656 for GPT-4V) and an Object HalBench hallucination rate of 10.3 percent versus GPT-4V's 13.6 percent.[^1][^11] Language coverage expanded to more than thirty languages, spanning Chinese, English, German, French, Spanish, Italian, Korean, Japanese, and a long tail of European and Asian languages defined in the model card's language list.[^11] The release also formalized streaming output and customizable system prompts as first-class features, and it was the first MiniCPM-V to be paired with an RLAIF-V alignment pass rather than the older RLHF-V approach.[^11] The 2.5 model card also documented LoRA fine-tuning on two NVIDIA V100 GPUs as a supported deployment path, which positioned the model as accessible to academic labs with modest hardware budgets.[^11]
The August 2024 release, paired with the formal technical report, dropped the Llama3 base in favor of Qwen2-7B, again landing at roughly 8B total parameters.[^2] MiniCPM-V 2.6 added two significant capabilities: multi-image reasoning (state-of-the-art on Mantis-Eval and BLINK at the model's scale, plus Mathverse mv and Sciverse mv) and full video understanding with temporal reasoning on Video-MME and Video-ChatGPT.[^2] OpenCompass scored the model at 65.2 on an updated eight-benchmark subset, with the authors claiming wins over GPT-4o mini, GPT-4V, Gemini 1.5 Pro, and Claude 3.5 Sonnet on single-image tasks.[^2] Token density was a marquee number: 1.8M-pixel images compressed to 640 visual tokens, around 75 percent fewer than the typical contemporary MLLM, enabling the demonstrated real-time video understanding on a stock iPad Pro.[^2] The OpenBMB team published a raw, unedited iPad Pro screen recording alongside the release to substantiate the on-device video understanding claim, a presentation pattern they repeated in later releases.[^2] In-context few-shot learning across multiple images, conversation and reasoning over image stacks, and chart and table understanding rounded out the capability bundle.[^2]
MiniCPM-o 2.6, released January 24, 2025, extended the architecture into a full speech-vision-text omnimodal model while staying at 8B parameters.[^3] The build was end-to-end across four pre-trained components: SigLIP-400M for vision, Whisper-medium-300M for audio understanding, ChatTTS-200M for speech generation, and Qwen2.5-7B as the language backbone.[^3] OpenBMB billed it as "a GPT-4o level MLLM for vision, speech, and multimodal live streaming on your phone."[^2] On OpenCompass the authors reported 70.2 average across the same eight-benchmark setup, ahead of GPT-4o-202405, Gemini 1.5 Pro, and Claude 3.5 Sonnet on single-image understanding for models below 25B parameters.[^3] On English speech recognition (LibriSpeech test-clean) the model achieved 4.4 percent WER, and on Chinese ASR (AISHELL-1) 1.6 percent CER, with the authors reporting results above GPT-4o-Realtime on audio understanding tasks.[^3] A new StreamingBench score of 66.0 measured the model's ability to process continuous video and audio streams without explicit user queries, ahead of GPT-4o-202408 and Claude 3.5 Sonnet on that benchmark.[^3] On real-time video specifically the StreamingBench sub-score was 79.9.[^3] Speech generation was rated on community ELO scales at 1088 semantic and 1163 acoustic.[^3] A novel Time-Division Multiplexing (TDM) mechanism handled the streaming omnimodal scheduling, and configurable audio system prompts let downstream applications swap voices in a relatively flexible way.[^3]
All MiniCPM-V models share a three-stage architecture: a frozen visual encoder (SigLIP SoViT-400m/14 in every version since 1.0), a compression layer based on a perceiver resampler that converts visual features into a small set of query tokens, and a decoder-only language model.[^1] The compression target is aggressive: in 1.0 and 2.0 each image slice is reduced to 64 query tokens, while MiniCPM-Llama3-V 2.5 uses 96 tokens per slice.[^1] By the 2.6 release the total visual token count for a full 1.8M-pixel image had settled at roughly 640 tokens, which the authors compare to thousands of tokens emitted by MLP-projector competitors at similar resolutions.[^2] The resampler is a single-layer cross-attention module: the small set of learned query tokens attends over the dense vision transformer feature grid and produces the fixed-length output that the LLM then consumes.[^1] Because the resampler runs once per slice rather than per generated text token, the cost of high-resolution visual context is paid up front, not amortized across every decode step.[^1]
The defining technical contribution from 2.0 onward is the adaptive visual encoding pipeline borrowed from LLaVA-UHD.[^7] Rather than resize an input image to a fixed square, the system computes an ideal slice count N as the ceiling of the input image area divided by the ViT's pre-training area, then searches row-column factorizations of N to pick the partition whose aspect ratio is closest to the source image.[^1] Each slice is independently resized to match the ViT's training area, with its 2D positional embeddings interpolated to the slice's true aspect ratio.[^1] The full original image is also resized and encoded as an extra slice so that the model sees both global context and local detail. Slice features are wrapped with <slice> and </slice> tokens, with newline characters demarcating rows so the spatial schema is preserved in the LLM's token stream.[^1] The net effect is that an arbitrarily shaped 1.8M-pixel image (for example a tall screenshot or wide receipt) can be encoded without distortion, which materially improves OCR and document-understanding accuracy.
The MiniCPM-V technical report describes a three-stage pre-training schedule.[^1] Stage 1 warms up the resampler at 224x224 resolution using roughly 200M image-caption pairs, with the ViT and LLM frozen. Stage 2 unfreezes the ViT and extends it to 448x448, training on another 200M samples to adapt the encoder to a higher fidelity regime. Stage 3 turns on the full adaptive encoding scheme and integrates dedicated OCR data, training both visual modules end to end at the 1.8M-pixel target.[^1] Supervised fine-tuning then runs in two parts: a recognition-focused phase using traditional VQA and captioning datasets, and a long-form interaction phase covering complex instructions across 36-plus languages, with roughly 2M curated samples in total.[^1] The two SFT phases are intentionally split: the first builds robust object, scene, and character recognition; the second teaches longer-form discourse and multi-turn reasoning, including responses that exceed the typical short-answer length of academic VQA datasets.[^1] The report frames this split as important for preventing the model from collapsing to terse, brittle responses, which is a common failure mode of MLLMs trained on captioning-style data alone.[^1]
For trust and hallucination control, MiniCPM-V uses RLAIF-V, a framework proposed in a companion paper by Tianyu Yu and colleagues (arXiv:2405.17220, May 27, 2024), later accepted as a CVPR 2025 highlight.[^10] RLAIF-V is built on two ideas. First, a divide-and-conquer feedback pipeline decomposes each candidate response into atomic claims (factored using a small text LLM such as Llama-3-8B), converts each claim to a yes/no question, and scores the questions with an open-source MLLM rather than calling GPT-4V or human annotators.[^10] Second, the resulting preference pairs are consumed by an online iterative form of Direct Preference Optimization (DPO) that mitigates the distribution-shift problem of vanilla DPO.[^10] The reported effect is large: at 7B scale RLAIF-V cuts object hallucination by roughly 80 percent and overall hallucination by roughly 34 percent, and at 12B scale a model trained against its own feedback can outperform GPT-4V on object hallucination benchmarks.[^10] For MiniCPM-V the practical signal is that MiniCPM-Llama3-V 2.5's 10.3 percent Object HalBench rate is below GPT-4V-1106's 13.6 percent at a fraction of the parameter count.[^1] The training set used to drive the DPO step is the publicly released RLAIF-V dataset on Hugging Face, which makes the alignment stage uniquely reproducible by comparison with most other open MLLM alignment pipelines that rely on private feedback corpora.[^11]
The technical report devotes a full section to on-device deployment, taking MiniCPM-Llama3-V 2.5 from a 16-17 GB FP16 footprint to a working 8B model running on a Qwen/Llama 3-class smartphone.[^1] Concretely, the authors used a Xiaomi 14 Pro powered by Qualcomm's Snapdragon 8 Gen 3 mobile platform as the reference target, with a vivo X100 Pro as a secondary device and an Apple MacBook Pro M1 included for comparison.[^1] A 4-bit Q4_K_M quantization via the GGML/llama.cpp toolchain reduced the memory footprint to roughly 5 GB. Sequential ViT/LLM loading cut peak memory further and brought image-processing time from 45.2 to 31.5 seconds. Native compilation lowered encoding latency from 50.5 to 17.0 seconds and improved decode throughput from 1.3 to 3.2 tokens/second. Automatic parameter tuning pushed throughput to 8.2 tokens/second on the Snapdragon target.[^1] Finally, porting the visual encoder to Qualcomm's QNN framework to run on the on-chip NPU reduced visual encoding from 3.7 seconds to 1.3 seconds, a roughly 150x speed-up over the unoptimized baseline.[^1][^11] The report concludes that throughput on the Xiaomi and vivo devices exceeds typical human reading speed, which is the authors' working threshold for "usable" deployment.[^1]
After that initial mobile demonstration, the inference story broadened. The MiniCPM-V 2.6 and MiniCPM-o 2.6 model cards document support for llama.cpp, Ollama, vLLM, int4 quantization at roughly 7 GB GPU memory, GGUF format in sixteen sizes, and integration with LLaMA-Factory for fine-tuning.[^2][^3] The model cards also document a Gradio WebUI for local interactive testing and the existence of an online demo space.[^2][^3] The combination of compact size, permissive licensing (Apache-2.0 code, free academic and registered commercial use of weights), and broad inference-runtime coverage is the principal reason MiniCPM-V is widely adopted as a default in open vision-language stacks.[^11]
| Version | Release | Total params | Vision encoder | LLM backbone | Key claim |
|---|---|---|---|---|---|
| MiniCPM-V 1.0 (OmniLMM-3B) | Jan/Feb 2024 | 3B | SigLIP-400M | MiniCPM-2.4B | Outperforms 9.6B Qwen-VL-Chat on MME/MMBench/MMMU at 3B[^5] |
| MiniCPM-V 2.0 | Apr 12, 2024 | 2.8B | SigLIP-400M | MiniCPM-2.4B | 1.8M-pixel adaptive encoding; matches Gemini Pro on scene text[^6] |
| MiniCPM-Llama3-V 2.5 | May 20, 2024 | 8B | SigLIP-400M | Llama3-8B-Instruct | First end-side MLLM at GPT-4V level on OpenCompass; 30+ languages[^11] |
| MiniCPM-V 2.6 | Aug 2024 | 8B | SigLIP-400M | Qwen2-7B | Multi-image + video; real-time video on iPad; 640 tokens per 1.8M-pixel image[^2] |
| MiniCPM-o 2.6 | Jan 24, 2025 | 8B | SigLIP-400M (+ Whisper-medium + ChatTTS) | Qwen2.5-7B | Omnimodal speech/vision/text with streaming; bilingual real-time speech[^3] |
The cross-row trend illustrates the project's strategy: keep the SigLIP-400M vision encoder constant, swap successively stronger language backbones (MiniCPM, Llama 3, Qwen2, Qwen2.5), and stack new modalities and alignment techniques without ballooning parameter counts beyond 8B.
The MiniCPM-V model family's applications cluster around scenarios where cloud-hosted MLLMs are impractical because of cost, latency, privacy, or connectivity. The model cards and technical report call out three main use cases.[^1][^2][^3]
Document and scene-text understanding is the most thoroughly evaluated. Reported OCRBench scores above 700 for both MiniCPM-Llama3-V 2.5 and MiniCPM-V 2.6, combined with native handling of 1.8M-pixel images at arbitrary aspect ratios, target tasks such as receipt parsing, ID-card and form OCR, screenshot question answering, and chart and table understanding, all on a smartphone or tablet.[^1][^2]
Multilingual visual assistance follows from the model's support for thirty-plus languages plus aggressive token compression: a 2.6 deployment can run with about 7 GB of GPU memory at int4 quantization, putting it within reach of mid-range consumer devices for translation-with-vision and accessibility tasks.[^2][^11]
Real-time multimodal interaction is the central pitch of MiniCPM-o 2.6, which accepts continuous video and audio streams independent of user queries and responds with bilingual speech.[^3] Together with the StreamingBench score of 66.0, that capability targets ambient assistants, video-call understanding, and on-device companion experiences without round-trips to a cloud server.[^3]
Outside the core MiniCPM-V repository the models have been packaged for Ollama, llama.cpp, and vLLM, making them a frequent default for developers building local vision language model prototypes that need both image and OCR capabilities under permissive licensing.[^2][^4]
Robotics and embodied AI form a fourth, more speculative application area. ModelBest's public marketing materials position the MiniCPM family as a candidate brain for smart home devices and robots, where the latency, privacy, and connectivity constraints of cloud MLLMs are particularly acute.[^8] Concrete benchmarks of MiniCPM-V or MiniCPM-o inside robotics stacks are sparser than the desktop OCR and ASR numbers, but the inference cost profile (10-plus tokens/second on Snapdragon 8 Gen 3 mobile, sub-7 GB int4 memory) is in the regime where running a multimodal policy on a battery-powered platform is plausible.[^1][^2]
A fifth application is education and accessibility. Because MiniCPM-V supports more than thirty languages and can run offline, the model has been used as a building block for offline study assistants, OCR-based reading aids for visually impaired users, and document translation pipelines in regions with intermittent connectivity.[^11] These uses are documented primarily in community forks and Hugging Face Spaces rather than in formal benchmarks, but they illustrate the practical reach of a freely downloadable 8B multimodal model.[^11]
MiniCPM-V's significance comes from being one of the first public demonstrations that strong multimodal capability does not require frontier-scale compute. The August 2024 technical report frames the contribution as evidence of a "rapidly decreasing" model size needed for usable GPT-4-V-class performance, paired with mobile silicon (the Snapdragon 8 Gen 3 generation specifically) that for the first time could host an 8B-parameter MLLM with NPU-accelerated visual encoding and 8-plus tokens/second decode.[^1] The combination undercut a widely held assumption that frontier multimodal models must be tens of billions of parameters and cloud-only.
A second contribution is the open release pattern. The code is Apache-2.0; weights are released on Hugging Face with free commercial use after a registration questionnaire.[^11] Together with the public RLAIF-V dataset and the LLaVA-UHD code, MiniCPM-V provided one of the most complete openly reproducible MLLM recipes in 2024, covering the vision encoder choice, the adaptive encoding scheme, the supervised fine-tuning mix, and the alignment pipeline.[^4][^10] Subsequent open multimodal projects, including later InternVL and Qwen2.5-VL revisions, share architectural family resemblance to the MiniCPM-V resampler-plus-SigLIP design, though those teams have published their own independent contributions.
Third, MiniCPM-V seeded a Chinese on-device AI ecosystem. ModelBest's pitch of "Little Powerhouses" engineered for smartphones, PCs, automotive systems, smart home devices, and even robots is now a competitive positioning against Gemini Nano, Phi-3 Vision, and proprietary on-device stacks from device OEMs.[^8] MIT Technology Review identified ModelBest as one of four Chinese AI startups worth watching beyond DeepSeek in its February 2025 coverage, noting the December 2024 funding round and the company's "tens of millions of dollars" valuation milestone.[^8]
Fourth, the project demonstrated that the alignment-data bottleneck for multimodal models could be cracked without proprietary feedback. The RLAIF-V framework, by using open-source MLLMs as labelers in an atomic-claim decomposition loop, lets a 7B model reduce object hallucination by roughly 80 percent without any reliance on GPT-4V or human annotators.[^10] That result has implications well beyond MiniCPM-V: it suggests that the gap between open and closed MLLMs on trustworthy behavior may be closable using fully open data and tooling, which informs the broader academic argument about open-source Multimodal Models.[^10]
Several caveats apply to the headline claims, drawn either from the technical report itself or from third-party model card commentary.[^1][^2]
Benchmark cherry-picking risk: the "GPT-4V level" claim is anchored to OpenCompass at a specific GPT-4V snapshot (1106) and an eleven-benchmark aggregate. On benchmarks not in that suite, particularly those involving long-horizon reasoning, complex spatial layouts, or video tasks beyond Video-MME, the gap to proprietary frontier models can be larger.[^1]
Token-density tradeoffs: aggressive resampler compression (640 tokens for a 1.8M-pixel image in 2.6) saves memory and decode time but can hurt fine-grained text recognition in dense documents or very small fonts, where MLP-projector models that emit thousands of visual tokens may still have an edge.[^2]
Mobile-NPU portability: the QNN-accelerated visual encoder demonstration is specifically against Snapdragon 8 Gen 3-class hardware, and the 8.2 tokens/second decode throughput on Xiaomi 14 Pro reflects a heavily optimized stack.[^1] On older or non-Qualcomm mobile chips the deployment story degrades substantially, a limitation acknowledged implicitly by the choice of reference hardware in the paper.
License nuance: although the code is Apache-2.0, commercial use of the weights requires registering a questionnaire with OpenBMB, which is less permissive than fully open licenses such as Apache-2.0 or MIT applied directly to weights.[^11]
Hallucination is reduced, not solved: even after RLAIF-V, Object HalBench rates of around 10 percent remain non-trivial, and out-of-distribution image domains (medical, scientific diagrams, low-resource languages) are not extensively covered in the released benchmarks.[^1][^10]
A high-profile dispute in mid-2024 around alleged training-data overlap with LLaVA derivatives also surfaced briefly in the open-source community, though the OpenBMB team published clarifications and the project remained widely used. That episode is not extensively documented in the formal academic record cited here and is therefore not described in detail.
Reproducibility constraints also affect external verification. The full training datasets are not all released, so independent replication of the OpenCompass and Object HalBench numbers reported in the technical report is harder than the open-weights and Apache-2.0 code license suggest at first glance.[^1] The RLAIF-V dataset is publicly available on Hugging Face, which closes that gap for the alignment stage specifically, but the multilingual pre-training mixture is described at a relatively high level in the paper rather than released as a single downloadable corpus.[^1][^10]
MiniCPM-V sits alongside several other open MLLM families targeting the 7-9B size class.
| Family | Vision backbone | LLM | Key feature | Comparison point |
|---|---|---|---|---|
| MiniCPM-V 2.6 | SigLIP-400M | Qwen2-7B | LLaVA-UHD adaptive encoding, RLAIF-V | OpenCompass 65.2, 640 visual tokens at 1.8M pixels[^2] |
| Qwen2.5-VL | Native dynamic ViT | Qwen2.5 | Native dynamic resolution, agent capabilities | Alibaba's primary open VLM line[^2] |
| InternVL | InternViT-6B | InternLM / Qwen | High-resolution multi-image | Larger total parameter counts at the top of the lineup[^4] |
| LLaVA (1.5/NeXT) | CLIP / SigLIP | Vicuna / Llama | Simple MLP projector | Reference baseline; higher token counts per image[^5] |
| DeepSeek-VL2 | Mixture-of-experts vision | DeepSeek-V2 / MoE LM | MoE multimodal scaling | Different efficiency strategy (MoE rather than compression)[^4] |
The closest peer in spirit is arguably Gemini Nano, which is similarly engineered for Edge AI deployment but ships as a closed system inside Android and Pixel devices, while MiniCPM-V is open-weights with a published technical report.[^8]