open model guide · september 2026
82 enterprise use cases and the open-weight model that solves each one. No closed APIs: everything here can be downloaded, audited and run under your control.
quick selector
Pick the type of task and your main constraint. The recommendation updates instantly; the 82 cases in detail are just below.
What kind of task?
What is your main constraint?
Qwen3.8-27BAPACHE 2.0
The newest of the Qwen generation and the strongest open writer you can run on one machine: 52 on the Artificial Analysis index, Apache 2.0, and native multilingual with solid Spanish.
the 82 cases
Filter by area or complexity. Each card gives the primary recommendation, the alternative and why. The badges under the model name say where you can run it without hosting it yourself: in Helmcode today, on the NaN.builders cluster, or soon in Helmcode.
11 open models cover the 82 cases. Click one to see only its own.
82 / 82 cases
Recommended open model
Gemma 4 26B
available in Helmcode
Low complexity and high volume, so the bill decides: $0.04 per task with 3.8B parameters active, and 256K of context holds a thread nobody wants to read twice.
Open alternative: DeepSeek V4 Flash, for threads that need the 1M window
Use case: Summarization →
Recommended open model
Gemma 4 26B
available in Helmcode
The whole company writes emails all day, and often not in one language, so coverage and cost per message decide: more than 35 languages out of the box, 140 in pretraining, and $0.04 per task with 3.8B parameters active. Multilingual by design, fine-tunable on how your team already writes, and on our flat rate.
Open alternative: Qwen3.8-27B, 27B dense on one machine
Use case: Content generation →
Recommended open model
Qwen3.8-Flash-Next
available on NaN.builders coming soon to Helmcode
1M of context takes a whole report, case file or book in one pass, with no chunking, and 56 on the Artificial Analysis index is what tells you it understood it. A window only says what fits. Its licence is not a free and open-source one, so read it.
Open alternative: DeepSeek V4 Flash, the same window at the market’s price floor
Use case: Summarization →
Recommended open model
Whisper large-v3-turbo
available in Helmcode · Whisper large-v3
Whisper large-v3-turbo transcribes the whole recording with no length cap, at 809M parameters and MIT, and V4 Flash turns the transcript into minutes.
Open alternative: Whisper large-v3 on our flat rate, 4.1% word error rate against turbo’s 4.6%
Use case: Voice & transcription →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
The context is already retrieved: extraction plus writing with precise citations. For confidential docs, self-hosting is the only valid option.
Open alternative: Qwen 3.6 35B
Use case: RAG over internal knowledge →
Recommended open model
Gemma 4 26B
available in Helmcode
Classification against clear criteria in milliseconds, on one GPU, over every email that comes in. Only 3.8B of its 25.2B parameters fire per token, which is what keeps the bill per message down.
Open alternative: DeepSeek V4 Flash
Use case: Classification & triage →
Recommended open model
Gemma 4 26B
available in Helmcode
Entity extraction (what, who, when) with reliable JSON, over everything the team writes in a day. At that volume cost decides, and 3.8B active parameters are the cheapest tier on the board.
Open alternative: DeepSeek V4 Flash
Use case: Autonomous agents →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Follow-ups should read as human, not automated. Same 27B dense footprint as the Qwen3.6-27B it replaces, and fourteen points higher on the index than it.
Open alternative: Qwen 3.6 35B, the MoE we serve on a flat rate
Use case: Autonomous agents →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Editing means respecting the author’s voice instead of rewriting it. Multilingual by design, 27B dense so it fits one machine, and cheap enough to run over everything the team writes.
Open alternative: GLM-5.3-Flash, or Qwen 3.6 35B on a flat rate
Use case: Content generation →
Recommended open model
Gemma 4 26B
available in Helmcode
No open model on this page covers more language for less: more than 35 languages out of the box, 140 in pretraining, Apache 2.0, and 3.8B parameters active per token, which is what makes translating a whole catalogue affordable.
Open alternative: Qwen3.8-27B for fine cultural adaptation, or V4 Flash at the highest volumes
Use case: Translation →
Recommended open model
qwen3-embedding + rerank
available in Helmcode
The forgotten half of every RAG stack: good embeddings plus reranking decide more than the generator model does.
Open alternative: DeepSeek V4 Flash (synthesis)
Use case: Semantic search & recommendation →
Recommended open model
qwen3-embedding + rerank
available in Helmcode
Reordering the retrieved top-k multiplies RAG precision at a minimal marginal cost.
Open alternative: Gemma 4 26B as cross-encoder
Use case: Semantic search & recommendation →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Comprehension plus rephrasing into user language. Ideal for manuals and policies that cannot leave for external APIs.
Open alternative: Qwen3.8-27B
Use case: Content generation →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
Data synthesis plus narrative: the best open reasoning following complex structured templates, and since 26 August the best of them costs a fifth of what it did.
Open alternative: GLM-5.3, which we serve on a flat rate
Use case: Content generation →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
One conclusion per slide, title as message. Wired to python-pptx or reveal.js: a full data-to-deck pipeline.
Open alternative: GLM-5.3 on our flat rate, or Qwen 3.6 35B on one machine
Use case: Content generation →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Precise language and consistent terminology across long documents. Internal content that should stay at home.
Open alternative: GLM-5.3-Flash, or Qwen 3.6 35B on a flat rate
Use case: Content generation →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
A 1-3h meeting is 50-150K tokens, so 1M of context holds a whole week of them. Whisper large-v3-turbo transcribes and V4 Flash writes the agenda and the minutes, both self-hosted.
Open alternative: GLM-5.3-Flash for meetings that end in a decision
Use case: Autonomous agents →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
RSS/scraping → batch summary and classification → dashboard. Cost per article is practically zero.
Open alternative: Gemma 4 26B
Use case: Summarization →
Recommended open model
Gemma 4 26B
available in Helmcode
One of the simplest tasks on the list: millisecond latency, on-premise, on one GPU, with 3.8B of its 25.2B parameters active per token.
Open alternative: DeepSeek V4 Flash
Use case: Classification & triage →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Task status, alerts and summaries. Milestones and owners of client projects never leave the internal network.
Open alternative: Qwen3.8-27B
Use case: Autonomous agents →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
The best open reasoning for synthesizing sources you have already gathered, with 1M of context to hold them. Pair it with your own scraping for the capture step.
Open alternative: GLM-5.3, or DeepSeek V4 Pro 0813
Use case: Research →
Recommended open model
GLM-5.3
available in Helmcode
Weights public since 25 August, and it matches Kimi K3 on half the weights: 88.2 on Terminal Bench 2.1 against K3’s 88.3, and 28.3 on Terminal Bench 3.0 against its 17.4, Z.ai’s own figures. We serve it on a flat rate, so the two things to plan for only apply if you run it yourself: 755 GB, about two 8-GPU nodes, and a licence that is MIT plus a clause aimed at $10bn model providers.
Open alternative: Kimi K3, or GLM-5.3-Flash when the budget is one node
Use case: Professional copilots →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
The open leader in agentic coding, and the lead grew on 26 August: Z.ai publishes 84.3 on Terminal Bench 2.1 and 63.4 on DeepSWE, its own figures, against 81.0 for GLM-5.2.
Open alternative: GLM-5.3, which we serve on a flat rate
Use case: Autonomous agents →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Reasons over the full diff to spot bugs, debt and security issues. With every pull request going through it, price decides: $0.11 per task against V4 Pro’s $0.27, and nothing leaves your infrastructure.
Open alternative: GLM-5.3 on our flat rate when the review is about security (84.5 on CyberGym, the best open figure), or DeepSeek V4 Pro 0813 for the hardest diffs
Use case: Professional copilots →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Edge-case coverage and coherent mocks, at the lowest price per task in the open field. DeepSeek publishes that Flash 0731 beats the V4 Pro preview on all nine of its agent benchmarks.
Open alternative: GLM-5.3-Flash, or DeepSeek V4 Pro 0813
Use case: Professional copilots →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
Hours-long multi-file refactors, where the model holds the repo and its own plan at once: 1M of context, and 328 GB of weights against the 1.5 TB GLM-5.2 needed for the same job.
Open alternative: GLM-5.3, which we serve on a flat rate
Use case: Code migration →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Text-to-SQL over your schema with a few-shot prompt. This is code, so the coding generation matters: Alibaba publishes 90.3 on LiveCodeBench v6. The schema is sensitive: better kept at home.
Open alternative: Qwen 3.6 35B, or V4 Flash
Use case: Professional copilots →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
Reads the code and writes the docs, holding coherence across modules in a large codebase. Z.ai publishes 84.3 on Terminal Bench 2.1 for that long multi-step work, with 1M of context.
Open alternative: GLM-5.3 on a flat rate, or Qwen3.8-27B on one machine
Use case: Content generation →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Resolves 80% of L1 queries by loading the whole KB into context. Employee and systems data: inside the network.
Open alternative: Gemma 4 26B
Use case: Customer support →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
Multi-step reasoning connecting unrelated events, with a method (5 Whys, Ishikawa). Confidential logs: self-host.
Open alternative: GLM-5.3, or DeepSeek V4 Pro 0813
Use case: Autonomous agents →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Reads the mockup and writes the component, with native vision and current-generation coding in the same 27B weights. It needed a 64-GPU cluster until August; now it needs one machine.
Open alternative: Qwen3.8-Flash-Next if the mockups justify a cluster
Use case: Professional copilots →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Training and test datasets at zero marginal cost. The MIT license places no restriction on how you use the outputs.
Open alternative: Qwen 3.6 35B
Use case: Synthetic data →
Recommended open model
Gemma 4 26B
available in Helmcode
Fine-tuned on your own categories of personal data, it runs at the edge of the pipeline before anything leaves.
Open alternative: DeepSeek V4 Flash
Use case: Anonymization →
Recommended open model
Whisper large-v3-turbo
available in Helmcode · Whisper large-v3
Pure transcription at call-centre volume, and the model everyone actually runs wins: 809M parameters, MIT, 99 languages and 4.6% word error rate on Artificial Analysis.
Open alternative: Whisper large-v3 when accuracy decides, on our flat rate
Use case: Voice & transcription →
Recommended open model
Chatterbox Multilingual
23 languages, Spanish among them, voice cloning from a few seconds of reference, MIT and 500M parameters. A full open voicebot: Whisper → LLM → Chatterbox. Kokoro only speaks English.
Open alternative: Kokoro, for English only and the smallest footprint
Use case: Voice & transcription →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Reads the document image and returns the fields as JSON in one pass, with up to 16,384 visual tokens per page to keep the small print readable, which is why the pick here is the LLM you already run. If OCR is the whole job, a specialist beats it: GLM-OCR is 0.9B under MIT and leads OmniDocBench v1.5 at 94.62.
Open alternative: GLM-OCR when the pipeline is pure OCR, Qwen3.8-Flash-Next for the highest accuracy tier
Use case: Document extraction →
Recommended open model
Qwen3.8-Flash-Next
available on NaN.builders coming soon to Helmcode
Visual QA, catalog tagging and asset verification. Alibaba publishes 88.5 against the 27B’s 85.9 on real-world perception, which is the axis this case turns on, and 6B active out of 125B keeps a whole catalogue affordable.
Open alternative: Qwen3.8-27B on one machine, or Gemma 4 26B on a flat rate
Use case: Visual inspection →
Recommended open model
Gemma 4 26B
available in Helmcode
Text and image in the same call, fine-tunable to your platform’s criteria for more consistency than zero-shot.
Open alternative: Qwen3.8-27B, Apache 2.0 and fine-tunable
Use case: Visual inspection →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Persuasive writing with frameworks (SPIN, Challenger) and a customer-centered narrative, and enough head to hold the argument across a long proposal.
Open alternative: GLM-5.3-Flash, or the full GLM-5.3 on a flat rate
Use case: Content generation →
Recommended open model
Gemma 4 26B
available in Helmcode
Fine-tuned on your conversion history, it beats zero-shot by a wide margin: the structural advantage of open source.
Open alternative: DeepSeek V4 Flash
Use case: Classification & triage →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Briefings from CRM activity. Amounts and clients under negotiation: commercial intelligence that must not leave.
Open alternative: Qwen3.8-27B
Use case: Autonomous agents →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Objection handling and personalization by profile. The sales script is confidential strategy: self-host.
Open alternative: Qwen 3.6 35B on a flat rate, or GLM-5.3-Flash for the hardest accounts
Use case: Content generation →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
The most solid open strategic reasoning that fits on one machine: 57 on the Artificial Analysis index and 328 GB of weights, against the 60 of GLM-5.3 at 755 GB and of Kimi K3 at 1.56 TB.
Open alternative: GLM-5.3, which we serve on a flat rate
Use case: Research →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Creativity, copywriting and adaptation to brand tone. At high volume, the saving over a closed API is substantial.
Open alternative: Qwen 3.6 35B on a flat rate at the highest volumes
Use case: Content generation →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Dozens of variations per campaign at zero marginal cost, and it reads the creative as well as the brief: native vision over image and video comes in the same weights.
Open alternative: Qwen 3.6 35B, or V4 Flash at the highest volumes
Use case: Content generation →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Structured batch generation across the whole catalog or blog. High volume, clear criteria: cost is what rules.
Open alternative: Qwen3.8-27B
Use case: Content generation →
Recommended open model
Gemma 4 26B
available in Helmcode
Thousands of opinions in batch with aspect-based analysis, at $0.04 per task. Fine-tuning it on your sector’s vocabulary is what sharpens the result, and that is only possible on open weights.
Open alternative: DeepSeek V4 Flash
Use case: Classification & triage →
Recommended open model
Gemma 4 26B
available in Helmcode
Your platform’s own criteria, fine-tuned, at a practically zero cost per message, and text and image go through the same call.
Open alternative: Qwen3.8-27B (fine-tune)
Use case: Classification & triage →
Recommended open model
Gemma 4 26B
available in Helmcode
Real-time multi-label, fine-tunable on your ticket history for business-specific precision.
Open alternative: DeepSeek V4 Flash
Use case: Classification & triage →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
A support chatbot with customer data under GDPR: in regulated sectors, self-hosting is not optional.
Open alternative: Qwen3.8-27B
Use case: Customer support →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Empathy plus firmness plus a concrete solution. Complaints contain personal data: process it at home.
Open alternative: Qwen 3.6 35B on a flat rate, or GLM-5.3-Flash
Use case: Customer support →
Recommended open model
Gemma 4 26B
available in Helmcode
Thematic classification and sentiment over thousands of open responses, at the lowest cost per response on the board, with the personal data never leaving EU infrastructure.
Open alternative: DeepSeek V4 Flash
Use case: Classification & triage →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Automatic closure of each ticket with a summary for the CRM. High volume, sound writing, one machine.
Open alternative: Qwen 3.6 35B, or V4 Flash at the highest volumes
Use case: Summarization →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Telling evidence from generic claims, without bias. Candidate data is GDPR territory: deploy locally.
Open alternative: GLM-5.3-Flash, or the full GLM-5.3 on a flat rate
Use case: Document review & QA →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
All the onboarding documentation fits in 1M of context, with no RAG and without internal policies leaving.
Open alternative: Gemma 4 26B
Use case: RAG over internal knowledge →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Constructive feedback with nuance. This data can end up in labor proceedings: self-host plus encryption.
Open alternative: GLM-5.3-Flash, or the full GLM-5.3 on a flat rate
Use case: Content generation →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
HR data is the most sensitive in the company: the conversations do not leave the corporate infrastructure.
Open alternative: Qwen3.8-27B
Use case: Professional copilots →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Persuasive writing with a defined structure, in native Spanish, even for confidential roles.
Open alternative: Qwen 3.6 35B
Use case: Content generation →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Instructional design adapted to the learner’s level. In high-volume e-learning, cost per module tends to zero.
Open alternative: Qwen 3.6 35B, the MoE we serve on a flat rate
Use case: Content generation →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Good tone and structure on one machine. Internal communications stay out of third-party APIs.
Open alternative: Gemma 4 26B, or Qwen 3.6 35B
Use case: Content generation →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Processes the invoice image directly and returns structured JSON, with up to 16,384 visual tokens per page against the 280 of the Gemma we serve. A scanned A4 is where the small print decides the amount. A 100% internal pipeline on one machine, and if that pipeline only ever reads documents, a specialist such as GLM-OCR (0.9B, MIT) is more efficient still.
Open alternative: Gemma 4 26B, which we serve on a flat rate
Use case: Document extraction →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
One index point behind V4 Pro at less than half the cost, with the same 1M of context for a full set of accounts and its notes. For listed companies or M&A, self-hosting is close to mandatory.
Open alternative: DeepSeek V4 Pro 0813, or GLM-5.3-Flash
Use case: Document analysis →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
Complex patterns over the full transaction history, and the 1M of context is the model’s own published number. In banking the data stays in, so MIT downloadable weights are what count.
Open alternative: GLM-5.3, or DeepSeek V4 Pro 0813
Use case: Classification & triage →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
High volume of short comparisons: statement lines against ledger entries, with the near-matches flagged for a person. Cost per token is what decides here.
Open alternative: Gemma 4 26B
Use case: Document review & QA →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Reads the accounts and writes the reasoning behind a limit, ratio by ratio, at $0.11 per file. A solvency judgement on a named company belongs self-hosted, and a person still signs it.
Open alternative: GLM-5.3-Flash for the files that go to committee
Use case: Document analysis →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
The same message has to escalate from a nudge to a formal notice without losing the client, so register weighs more than raw reasoning. 27B dense weights under Apache 2.0.
Open alternative: Qwen 3.6 35B, the MoE we serve on a flat rate
Use case: Content generation →
Recommended open model
Gemma 4 26B
available in Helmcode
Judgement against clear criteria, which is what this size is for: this dinner is over the per diem, this taxi has no attendees. The policy travels in the prompt and changes without retraining anything, at $0.04 per task.
Open alternative: DeepSeek V4 Flash
Use case: Document review & QA →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
The highest open level of nuance comprehension you can deploy, with MIT license and model card published: auditable end to end.
Open alternative: GLM-5.3, or DeepSeek V4 Pro 0813
Use case: Document review & QA →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
With the AI Act applying in general since 2 August 2026, an open, traceable stack on EU infrastructure gives your assessment the evidence it needs.
Open alternative: GLM-5.3, which we serve on a flat rate
Use case: Document review & QA →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Tagging and triage of case files by matter, jurisdiction and urgency, without a single page leaving the firm.
Open alternative: Qwen 3.6 35B on a flat rate, or Gemma 4 26B
Use case: Classification & triage →
Recommended open model
Kimi K3
Thousands of documents that only mean something read together, so reasoning across them matters more than window size: K3 leads every open model on AA-LCR at 82.7%.
Open alternative: GLM-5.3 on our flat rate, the same index score on half the weights, or DeepSeek V4 Flash on hardware you already have
Use case: Document analysis →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
Reasoning that has to hold a chain of citations without inventing one. Ground it with RAG on your own database of rulings, never on the model memory.
Open alternative: GLM-5.3, or DeepSeek V4 Pro 0813
Use case: Research →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Generation rather than analysis: assembling a first draft out of clauses legal has already approved, so the team edits instead of starting from nothing.
Open alternative: GLM-5.3-Flash, or Qwen 3.6 35B on a flat rate
Use case: Content generation →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
Watching what changed in a regulation and which internal policies it touches. A different job from checking compliance: this one runs before anybody is out of it.
Open alternative: DeepSeek V4 Flash
Use case: Autonomous agents →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Comparing terms and flagging expirations. Agreed prices are confidential commercial information.
Open alternative: GLM-5.3-Flash (every contract at once, 1M ctx)
Use case: Autonomous agents →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
Structured reasoning over sources you have already gathered. For web capture, combine it with your own scraping.
Open alternative: DeepSeek V4 Flash
Use case: Research →
Recommended open model
Qwen3.8-27B
coming soon to Helmcode
Hundreds of requirements answered one by one, each traceable to a document you already hold. The bid stays confidential until the envelope is opened.
Open alternative: GLM-5.3-Flash, or the full GLM-5.3 on a flat rate
Use case: Content generation →
Recommended open model
GLM-5.3-Flash
available on NaN.builders coming soon to Helmcode
The same two hundred questions arrive from every client with the wording changed. Answered from your own evidence base, with a person signing it off.
Open alternative: DeepSeek V4 Flash
Use case: Document review & QA →
Recommended open model
DeepSeek V4 Flash 0731
available in Helmcode
Thousands of short events a day, each needing a route: delay, damage, wrong address or nothing at all. Volume picks the model.
Open alternative: Gemma 4 26B
Use case: Classification & triage →
Recommended open model
Qwen3.8-Flash-Next
available on NaN.builders coming soon to Helmcode
Reads the picture off the line, or the clip, and writes the defect report against your own criteria. Alibaba puts it at 88.5 on real-world perception, its best published figure of any model it compares itself against, and a missed defect costs more than the inference does.
Open alternative: Qwen3.8-27B on one machine, or Gemma 4 26B on a flat rate
Use case: Visual inspection →
Recommended open model
FLUX.2 [klein] 9B
available in Helmcode available on NaN.builders
Dozens of pieces a week with no stock licence and no agency round: 9B parameters, about 29 GB of VRAM, and a generation in under a second. On our flat rate, and on the NaN.builders cluster. Its licence is not a free and open-source one, so read it before it goes into a commercial pipeline.
Open alternative: FLUX.2 [klein] 4B, the Apache 2.0 route on one consumer GPU, or Qwen-Image-2512
Use case: Content generation →
Recommended open model
FLUX.2 [klein] 9B
available in Helmcode available on NaN.builders
Backgrounds, formats and versions of the same shot: editing takes several reference images and runs on the same weights that generate, so a catalogue is one model and not two pipelines. On our flat rate.
Open alternative: Qwen-Image-2512, Apache 2.0, when the piece leans on text inside the image
Use case: Content generation →
In production 99.5% of our tokens go through open models, and not on the easy tasks: classifying, extracting, summarizing, drafting, answering and reasoning over documents nobody else has read. What is left is a narrow set of genuinely frontier problems, and the answer to those is to route them to the big model, not to pay frontier prices for the other 99.5%.
go deeper
The 82 cases above roll up into 20 canonical use cases, and every card links to its own. Each has its own page with architecture, FAQ and deployment in detail.
get started
Pick a case from the list, point the same code at our endpoint and read both outputs side by side.
The API you already call. One afternoon. Nothing to migrate.
A process with volume and stable criteria can become a small, specialized model with the weights in your name, running on far less hardware. model_specialization →
// cookies
We only use strictly necessary cookies to run the site. No analytics, no advertising, ever — see our Cookie Policy.
// preferences