Specialized LLMs: A Guide to Bespoke Labs' Open Models

Purpose: Bespoke-MiniChart-7B is a vision-language model (VLM) designed to answer questions about charts and graphs. In other words, given a chart image and a question (e.g. “What trend does this line chart show?”), it generates an answer. This task requires both precise visual perception (to read axes, labels, and data points) and logical reasoning (to interpret trends and comparisons). Chart QA is challenging because charts can vary widely in style and data, and questions often require multi-step “chain of thought” reasoning.

Model details: This model is branded 7B and is built on the Qwen2.5-VL-7B-Instruct foundation; its hosting metadata reports 8B total parameters. Bespoke Labs chose this base VLM for its strong out-of-the-box performance (^[4]). They then fine-tuned it on a large synthetic chart-QA dataset of about 1 million examples (built from 13K chart images and many question-answer pairs, each with chain-of-thought). Training was done in three stages (supervised fine-tuning, rejection-sampled CoT collection, and a second fine-tune), followed by Direct Preference Optimization (DPO) to improve reasoning paths (^[5]) (^[4]). The Bespoke-MiniChart-7B weights are publicly available under the non-commercial CC BY-NC 4.0 license. Its model card reports state-of-the-art chart-question-answering results among 7-billion-parameter models and comparisons with larger closed models on seven public benchmarks (^[6]) (^[1]).

• Key points: It is a Vision-Language Model (VLM) – i.e. it takes both an image and text as input (^[6]). It was trained on carefully curated synthetic chart data, which made it robust to real-world charts. According to Bespoke, “data curation matters” – the curated dataset significantly improved the model’s performance on out-of-distribution charts (^[5]). Also, they found that simple fine-tuning (SFT) was not enough; using DPO to optimize answer preferences boosted its chain-of-thought reasoning (^[5]).

• Performance: Bespoke reports that, despite its 7B branding (with hosting metadata reporting 8B total parameters), MiniChart matches larger models including Gemini-1.5-Pro and Claude-3.5 across seven benchmarks. The publicly available weights are licensed CC BY-NC 4.0, which restricts commercial use (^[6]) (^[1]).

Purpose: Bespoke-MiniCheck-7B is specialized for factual groundedness: given a reference document and a sentence (claim), it decides whether the claim is supported by the document. In other words, it’s a built-in fact-checker. This is crucial for systems like retrieval-augmented generation (RAG), where you want the final answer to be faithful to source content. MiniCheck focuses on grounded factuality: verifying claims against context (^[7]).

Model details: MiniCheck is fine-tuned from the internlm/internlm2_5-7b-chat base model. Bespoke Labs fine-tuned it on a small curated dataset (35K examples total) specifically constructed for factuality checking; Meta Llama 3.1 405B Instruct generated the synthetic portion of that training data (^[8]). The training mix included 21K examples from the ANLI textual entailment dataset and 14K synthetic examples generated according to the “MiniCheck” methodology (claims paired with documents) (^[8]). This 35K-example figure describes Bespoke’s additional fine-tuning; MiniCheck retains the substantial pretraining of its InternLM2.5 base model.

• Key points: MiniCheck is a binary judgment model: it answers “Yes/No” on whether the document supports the claim (^[9]) (^[10]). It is optimized to be concise and factual. Despite its simplicity, Bespoke claims it is extremely effective: their research team reports that “Bespoke-MiniCheck-7B is lightweight and outperforms all big foundation models including GPT-4o (now succeeded by GPT-5.2) and Mistral-Large 2 for this specialized task” (^[2]). In fact, MiniCheck “tops the LLM AggreFact leaderboard” (a community benchmark for factuality) and achieves state-of-the-art fact-checking accuracy with only 7B parameters (^[7]).

• Usage: The model takes a long text document (up to 32K tokens) and a claim sentence. It then outputs “Yes” (if the document supports it) or “No”. It can check multi-sentence claims by evaluating each sentence separately (^[11]). Bespoke releases the model weights under CC BY-NC 4.0; its model card says that commercial licensing is available by contacting the company. OLLAMA’s hosting page emphasizes: “ [Bespoke-MiniCheck] is a state-of-the-art fact-checking model… and despite its small size, it is SOTA for its domain” (^[12]).

Purpose: The OpenThinker series are general reasoning LLMs (one with 7B and another with 32B parameters) fine-tuned for complex math, code, and science questions. Their weights, training data, and associated code are publicly available. Bespoke reports an inadvertent political-decensoring effect after fine-tuning on its reasoning dataset; this does not mean the models were intentionally stripped of general safety protections or designed to disregard safety rules. The models are intended for structured reasoning tasks with publicly available training artifacts.

Model details: Both OpenThinker-7B and -32B are built by fine-tuning Qwen2.5-7B-Instruct and Qwen2.5-32B-Instruct (open-source bases) on OpenThoughts-114k, a newly curated reasoning dataset. This dataset (released by the same consortium behind Bespoke) contains 114,000 high-quality examples spanning math puzzles, coding problems, and science questions (^[13]). The dataset was derived by distilling DeepSeek-R1. Bespoke says it contained no political questions intended to produce decensoring, to the best of the team's knowledge.

Bespoke reports that fine-tuning Qwen on this dataset changed how the resulting models responded to politically sensitive prompts, even though the dataset was not designed around political questions (^[14]) (^[15]) (^[16]). The 7B model was announced in January 2025 and the 32B in February 2025. Both are fully open: Hugging Face hosts their weights and datasets, and Bespoke provides all training and evaluation code. Since the initial release, the OpenThinker line has evolved rapidly: OpenThinker2 (7B and 32B) launched in April 2025 using the much larger OpenThoughts2-1M dataset, with OpenThinker2-32B becoming the first open-data model to match DeepSeek-R1-Distill-32B on standard reasoning benchmarks (^[17]). OpenThinker3-7B followed in June 2025, trained on the OpenThoughts3-1.2M dataset (850K math, 250K code, and 100K science problems), achieving state-of-the-art results among open-data 7B reasoning models (^[18]). Most recently, in December 2025, the project expanded into agentic AI with OpenThinker-Agent-v1, a model post-trained from Qwen3-8B for terminal and software engineering tasks, ranking as the best model of its size on Terminal-Bench 2.0 (^[19]).

• Key points: OpenThinker models specialize in “reasoning” tasks, often requiring multi-step chain-of-thought answers. They are benchmarked on math (MATH500, AIME24), coding puzzles, logical reasoning, etc. In published results, OpenThinker-32B was noted to beat DeepSeek’s corresponding 32B model on several benchmarks (^[15]). The 7B OpenThinker surpasses its own predecessors (e.g. Bespoke’s earlier Stratos-7B) due to the larger 114K training data (^[20]). Unlike many commercial LLMs, these are intended for transparency and research – for example, Bespoke explicitly lists that “our model weights, datasets, data generation code, and training code are all publicly available” (^[21]).

• Decensoring: Bespoke Labs describes OpenThinker as “decensored,” referring to how it responded to politically charged prompts in the company’s testing. Bespoke Labs reported differences between OpenThinker and the Qwen base models used in that evaluation (^[22]). The team describes this as a side effect of focusing only on neutral reasoning tasks in training data – they “created an open reasoning dataset with no political questions, and the fine-tuned models appear to be decensored (but still aligned against unsafe content)” (^[23]) (^[16]).

• Availability: The OpenThinker-7B model card lists an Apache-2.0 license, and Bespoke Labs’ February 2025 post reported that the 7B and 32B models were available through Hugging Face, Ollama, and its playground. The article does not independently verify later usage figures or conference-acceptance claims (^[24]).

In January 2025, Bespoke Labs also released the Bespoke-Stratos series (7B and 32B), reasoning models distilled from DeepSeek-R1 using Berkeley NovaSky's Sky-T1 data pipeline. Stratos-32B was trained on just 17,000 curated examples (a mix of programming, mathematics, and science/puzzle data) yet outperforms Sky-T1 and OpenAI's o1-preview on reasoning benchmarks, nearly matching DeepSeek-R1-Distill-Qwen-32B while being trained on 47x fewer examples (^[25]). The Stratos models served as a precursor to OpenThinker, demonstrating Bespoke's thesis that careful data curation matters more than sheer dataset scale.

Bespoke Labs' strategic focus has shifted since the model releases described above. On July 6, 2026, the company announced $40 million in total funding: an earlier $8.25 million seed round led by 8VC, with participation from Google DeepMind chief scientist Jeff Dean, plus a new $31.75 million Series A led by Wing VC, joined by Mayfield, The House Fund, and individual angel investors from Anthropic, OpenAI, and Meta (^[26]) (^[27]). Bespoke Labs was founded in 2024 by Mahesh Sathiamoorthy and Alex Dimakis, and it now describes its mission as building the reinforcement learning environments and infrastructure that let frontier labs and enterprises train, evaluate, and improve long-horizon AI agents for production use, rather than releasing standalone specialized checkpoints (^[26]).

The company frames this pivot around a specific bottleneck: agents remain unreliable over long, autonomous task horizons, and the training environment, not the base model, is what determines whether an agent can be trusted in production. As Bespoke put it, the environment an agent learns in "is the only important component that is not going to be democratized and thus will ultimately determine whether the agent is reliable enough to be trusted in production" (^[26]). New funds go toward expanding the research team and scaling infrastructure that simulates realistic settings, for example employee workstation sandboxes for productivity agents or simulated GitHub repositories for coding agents, then optimizes model behavior in those settings using tools such as GEPA, an open-source, genetic-algorithm-based optimizer for agent prompts and policies (^[27]).

This shift places Bespoke alongside a broader industry move away from pure pretraining scale and toward reinforcement-learning-based post-training as the next competitive axis for reliable agents. As company-reported evidence of traction from its earlier open-source work, Bespoke says OpenThoughts has been downloaded hundreds of thousands of times and has powered work at Meta, Amazon, and AI2; it also says Anthropic, OpenAI, and Google DeepMind use Terminal-Bench to showcase frontier models’ agentic abilities (^[26]). For a company that made its name on small, task-specialized language models, the emphasis on environments over checkpoints signals that data and simulation curation, not parameter count, remains its core differentiator.

Though Bespoke's publicly released models differ in license as well as in purpose, design, and training, they differ fundamentally in the following ways:

• Task specialization: Each model targets a narrow domain. MiniChart-7B is a Vision-Language model for chart interpretation; MiniCheck-7B is a binary fact-checker working over documents; OpenThinker models are general text-only reasoners for math/code puzzles. They do not try to do everything. In contrast, many other open models (like LLaMA, Falcon, Mistral) are general-purpose and not optimized for these specific tasks.

• Input/output format: MiniChart takes images + text and outputs free-form answers (often multi-sentence reasoning). MiniCheck takes (very long) text plus a claim, and outputs simply “Yes” or “No”. OpenThinker takes a textual question and produces a step-by-step solution followed by an answer. This reflects their different goals: chart Q&A vs. binary verification vs. chain-of-thought reasoning.

• Architecture and base models: MiniChart-7B was built on a Vision-Language foundation (Qwen2.5-VL-7B). MiniCheck-7B is a pure text LLM fine-tuned from internlm/internlm2_5-7b-chat; Llama 3.1 405B generated its synthetic training examples. OpenThinker models started from Qwen2.5 text-only models (7B and 32B). In other words, Bespoke chose different pre-trained backbones suited to each task before fine-tuning.

• Training data and methods: A common thread is that Bespoke heavily curated synthetic data. MiniChart was trained on ~1 million QA examples with chain-of-thought reasoning about charts (^[28]). MiniCheck was trained on a small mixture of human-generated entailment examples (ANLI) plus Bespoke’s own synthetic claims (^[8]). OpenThinker’s training was the OpenThoughts-114k dataset distilled from other models and verified. Importantly, Bespoke often uses multi-stage training (e.g. SFT → rejection sampling for more CoTs → DPO for MiniChart (^[5]) (^[4])) to squeeze extra performance out of the data. This contrasts with some other open LLM projects that focus on scaling up raw compute or data size rather than careful synthetic curation.

• Performance claims: Despite being open and smaller, these models claim to outperform many larger systems on their tasks. Bespoke emphasizes that MiniChart-7B “sets a new state-of-the-art” and beats giants like Claude and Gemini on chart QA (^[6]) (^[1]). MiniCheck-7B “outperforms all big foundation models including GPT-4o (now succeeded by GPT-5.2) and Mistral-Large 2” at factuality checking (^[2]). While OpenThinker 7B/32B may not match GPT-4 on all benchmarks, they score competitively on math and code problems and have the advantage of open reproducibility. The key takeaway is: by focusing on one domain and fine-tuning carefully, a 7B Bespoke model can rival or exceed a 35B+ closed model for that task.

• Availability and licensing: Bespoke publicly releases model artifacts, but terms vary by model. The OpenThinker-7B model card lists Apache-2.0, while the MiniChart and MiniCheck model cards list CC BY-NC 4.0 and direct prospective commercial users to contact Bespoke. MiniChart and MiniCheck are therefore publicly available, non-commercially licensed model weights—not open-source releases in the conventional licensing sense (^[21]).

In summary, Bespoke Labs' publicly released models are task-focused tools rather than clones of ChatGPT. Each model is different: a MiniChart VLM for visual Q&A, a MiniCheck text model for factual verification, Stratos models for efficient reasoning distillation, and the expanding OpenThinker family for logical problem-solving and now agentic tasks. They differ from each other in inputs, training data, and objectives; and they differ from other LLMs by being smaller, highly specialized, critically evaluated on public benchmarks, and released with open datasets and code. The rapid evolution from OpenThinker (January 2025) through OpenThinker3 (June 2025) and into OpenThinker-Agent (December 2025) demonstrates how quickly the field is advancing — and Bespoke's commitment to keeping that progress open. Together, these models illustrate a curation-focused approach to LLM development. Their published model cards and blog posts document the projects' licenses and vendor-reported evaluations, which should be read as model-specific claims rather than independent rankings (^[1]) (^[2]) (^[21]).

Sources: Information is drawn from Bespoke Labs' official blog posts, research papers, and model card documentation (^[1]) (^[6]) (^[2]) (^[21]), (^[15]), the OpenThoughts project site (^[17]), and the Bespoke Stratos blog post (^[25]). The models themselves are publicly available on Hugging Face and integrated into platforms like Ollama, reinforcing their open availability.