// models
Open models.
Frontier results.
The open frontier already covers the work enterprises actually run, retrieval, code, agents, extraction. Benchmarked against the closed labs, proven in production, served only from the EU.
// our models
Models we serve at Helmcode.
The complete catalogue on the Shared cluster: 9 models behind one OpenAI-compatible API, all of them open-weight, all served from the EU. Model IDs are shown as-is, copy them straight into your code.
Language models
glm5.3 GLM-5.3 (custom) 744B MoE · 40B active · NVFP4 · 1M ctx · add-on Long-horizon agentic coding and hard maths. Multimodal: text, images, video and audio. Sold per key, €150 a month. glm5.3-flash MIT 320B MoE · 18B active · NVFP4 · 1M ctx Agentic coding and long context, on your plan. Image and video input. deepseek-v4-flash MIT 552B MoE · 8B-16B active · NVFP4 · 1M ctx DeepSeek V4.1 Flash. Reasoning, agents and long-context analysis. Native tool calling. qwen3.6 Apache 2.0 35B MoE · 3B active · NVFP4 · 256K ctx High-volume RAG, classification and code. Multimodal input. gemma4 Gemma 26B MoE · 4B active · NVFP4 · 256K ctx Efficient assistants, documents and scans. Multimodal input. Embeddings & reranking
qwen3-embedding Apache 2.0 8B · 4096 dims · Float32 · MMTEB 70.58 The retrieval layer of a RAG pipeline, 100+ languages. rerank Apache 2.0 Qwen3-Reranker-8B · BF16 · /v1/rerank Reorders the retrieved top-k by real relevance. Speech
whisper MIT 99+ languages · 3.2% WER (Spanish) Speech to text for calls, meetings and media. kokoro Apache 2.0 82M params · <1s latency · 67 voices Text to speech for live agents and IVR, Spanish included. Every model here runs on Helmcode today. On Dedicated and On-premise plans we also serve custom or fine-tuned open-weight models on hardware reserved for you.
Full model reference in the docs// the 80 / 20
Open models cover 80% of enterprise inference.
The work that runs your business, RAG, classification, code generation, internal assistants, is solved today by open-weight models. The 20% where you truly need a frontier closed model is narrower than it looks.
Bleeding-edge reasoning at the very limit of capability. Real, but specific, and rarely the workload a regulated team needs to keep in-house. We are honest about that boundary instead of pretending it doesn't exist.
// benchmark · artificial analysis
Frontier-class intelligence, open-weight prices.
The Artificial Analysis Intelligence Index is a composite of ten hard evaluations, agentic work, coding, long-context and scientific reasoning: Terminal-Bench, SciCode, Humanity's Last Exam, GDPval and more. The best open-weight model sits a step below the closed frontier, at a fraction of the price.
GLM-5.3 leads the models you can actually download, at 45, one point ahead of Kimi K3 (44), against 58 for Claude Opus 5.5, the closed leader. We serve GLM-5.3 as a per-key add-on, so the highest open score on the board is inside the platform. Behind it sit the models your plan covers: DeepSeek V4.1 Flash (39), Qwen3.6 35B (18) and Gemma 4 26B (17). V4.1 Flash scores 39 at $0.15 / $0.60 per million tokens against Claude Opus 5.5's $4.00 / $20.00, that's ~27× cheaper to read and ~33× cheaper to write, for 67% of the index leader's intelligence.
Source: artificialanalysis.ai · Intelligence Index v4.3 · October 2026 · composite of 10 evaluations · index rounded to whole points · price = first-party API, per 1M tokens (input · output). Gemini 4 Argon is at Google's introductory $2/$10, which it says will double, and on 1 October it was not yet open to the public API. The two DeepSeek figures are the off-peak rate and double during peak hours, 01:00-04:00 and 06:00-10:00 UTC, under the peak/off-peak rate card DeepSeek has used since 16 August 2026. Rows marked on helmcode, GLM-5.3, DeepSeek V4.1 Flash, Qwen3.6 35B and Gemma 4 26B, are the ones we serve, with GLM-5.3 sold per key rather than covered by the plan. GLM-5.3 ships under Z.ai's own GLM-5.3 licence rather than MIT: the MIT text plus one clause, a Z.ai security review for any model-as-a-service business above $10bn of revenue. On Helmcode it is multimodal: text, images, video and audio. Kimi K3 and DeepSeek V4 Pro 0813 are open-weight but not on the platform: K3 needs roughly 64 GPUs of the H100 or B200 class to serve, against about 16 for GLM-5.3, and its weights ship under Moonshot's own K3 licence rather than an OSI-approved one.
// proven in production
The production numbers.
On the infrastructure we run, effectively all inference already runs on open models, led by DeepSeek V4-Flash.
2.8T
Tokens in production
cumulative
43.7%
Run on deepseek-v4-flash
the open workhorse
>99.9%
Of tokens on open models
every token served
// the lineup
What actually runs the 80%.
The five language models in the catalogue, ordered by their share of every token we serve. The remaining ~16% runs on community-side models and embeddings, all on OpenData.
deepseek-v4-flash 552B MoE · 1M ctx 43.7% Reasoning, agents, long-context qwen3.6 35B MoE · 256K ctx 22.6% High-volume RAG, classification, code glm5.3-flash 320B MoE · 1M ctx 14.7% Coding agents, long context glm5.3 744B MoE · 1M ctx · add-on 2.1% Long-horizon agentic coding, per key gemma4 26B MoE · 256K ctx 1.2% Efficient assistants, document work Plus embeddings & reranking (qwen3-embedding, rerank) and speech (kokoro, whisper), nine models on one API. Full model reference →
// models faq
Open models, answered.
The questions everyone asks before trusting open models in production.
Are open models actually good enough?
For the work enterprises run day to day, yes. On Artificial Analysis’ Intelligence Index, the leading open-weight model, GLM-5.3, scores 45 against 58 for the closed leader, Claude Opus 5.5, and we serve it. DeepSeek V4.1 Flash, which your plan covers, delivers 67% of the leader’s score for cents per million tokens. In production, more than 99.9% of the tokens we serve flow through open models. The gap that remains is a narrow set of frontier tasks most teams never hit.
Which model should I use?
More production tokens run on deepseek-v4-flash (today DeepSeek V4.1 Flash) than on any other model: reasoning, agents, 1M-token context. For high-volume RAG, classification and code, Qwen 3.6 is the fastest path and unlimited on every plan. For coding agents, GLM-5.3 Flash; for long-horizon work, the GLM-5.3 add-on. Image input: Qwen 3.6, Gemma 4, GLM-5.3 Flash and the GLM-5.3 add-on. Same API, just change the model id.
What about the 20% that genuinely needs a frontier model?
It exists, and it is more specific than most assume, frontier-only reasoning at the very edge of capability. Helmcode is honest about that boundary: we cover the 80% that runs your business, privately and at a flat rate, not the last mile of the leaderboard.
How current are these benchmarks?
Figures are Artificial Analysis Intelligence Index v4.3, read 1 October 2026: the same independent evaluation for open and closed models. Benchmarks move every release, so treat them as directional. What does not move is where your data is processed: always the EU, always zero logs.
Can I run a model that is not listed?
On Dedicated and On-premise plans, yes, custom or fine-tuned open-weight models on hardware reserved for you. The Shared cluster serves the curated lineup above.
// get started
START BURNING TOKENS
Skip the AI infra work. Deploy your first private inference endpoint today.
Flat rate. EU data. OpenAI API compatible.