NERIN
SHEET 04

Equipment schedule

21 marks

Open weights, run on our own accelerators. Every rate below is derived from the cost model on sheet 02 — none of it is a list price we chose.

MarkModelParametersLicenceContextReplicaThroughputCostListPer
H-01CodeSoftware engineering agents
M-01Qwen3-Coder-480B-A35BDefault planner and editor for the code agent.480B / 35B actApache-2.0262.144k4 GPU3.2k tok/s$0.531$1.061M tok
M-02DeepSeek-V3.2Long-horizon reasoning; used for repository-wide refactors.685B / 37B actMIT163.84k8 GPU5.4k tok/s$0.629$1.261M tok
M-03Llama-4-MaverickWhole-codebase context without chunking.400B / 17B actLlama 4 Community1,000k8 GPU6.1k tok/s$0.557$1.111M tok
M-04Kimi-K2 (1T-A32B)Strongest open agentic tool-use; reserved for hard runs.1000B / 32B actModified MIT131.072k16 GPU5.2k tok/s$1.31$2.611M tok
M-05GLM-4.6Balanced cost/quality workhorse.355B / 32B actMIT200k8 GPU4.4k tok/s$0.772$1.541M tok
M-06Qwen3-32BDense, latency-sensitive edits and completions.32BApache-2.0131.072k2 GPU2.4k tok/s$0.354$0.7071M tok
M-07Devstral-SmallSingle-GPU tier for cheap loops and CI checks.24BApache-2.0131.072k1 GPU1.9k tok/s$0.223$0.4471M tok
H-02Biological designProtein language models
M-08ESM-C 6BRepresentation model — embeddings and variant effect scoring.6BEvolutionaryScale Community1 GPU15.0k res/s$0.028$0.0571M res
M-09ESM3-open 1.4BGenerative PLM — sequence, structure and function conditioning.1.4BEvolutionaryScale Community1 GPU4.0k res/s$0.106$0.2121M res
M-10ProGen2-xlargeFamily-conditioned de novo sequence generation.6.4BApache-2.01 GPU1.5k res/s$0.283$0.5661M res
M-11Boltz-2Co-folding and binding affinity in one pass.1.2BMIT1 GPU0.05 struct/s$8.49$16.981k struct
M-12RFdiffusionBackbone design for binders and scaffolds.60MBSD-3-Clause1 GPU0.2 backbone/s$2.12$4.241k backbone
H-03VisionVision-language models
M-13Qwen3-VL-235B-A22BFrontier-class open VLM; documents, video, GUI grounding.235B / 22B actApache-2.0262.144k8 GPU3.9k tok/s$0.871$1.741M tok
M-14InternVL3-78BHigh-resolution tiling for dense diagrams and scans.78BMIT32.768k4 GPU2.2k tok/s$0.772$1.541M tok
M-15Molmo-72BPointing and spatial grounding with open training data.72BApache-2.032.768k4 GPU2.0k tok/s$0.849$1.701M tok
M-16SAM 3Promptable segmentation and tracking.900MSAM Licence1 GPU30 img/s$0.014$0.0281k img
M-17dots.ocrLayout-preserving document parsing to structured text.1.7BMIT1 GPU2.5 page/s$0.170$0.3401k page
H-04KnowledgeRetrieval, embedding, synthesis
M-18Qwen3-Embedding-8BMultilingual retrieval embeddings.8BApache-2.032.768k1 GPU96.0k tok/s$0.004$0.0091M tok
M-19BGE-M3Dense, sparse and multi-vector retrieval in one pass.570MMIT8.192k1 GPU210.0k tok/s$0.002$0.0041M tok
M-20Qwen3-Reranker-4BCross-encoder rerank over retrieved candidates.4BApache-2.032.768k1 GPU64.0k tok/s$0.007$0.0131M tok
M-21Llama-4-ScoutTen-million-token context for whole-corpus synthesis.109B / 17B actLlama 4 Community10,000k4 GPU3.4k tok/s$0.499$0.9991M tok

21 models in service. Input billed at 25% of the listed output rate. Licences are as published by each model's authors — check them before commercial use. Prices are derived from the cost model in note 02.1 and change when it changes.

ISSUE

Issue and revision

Bring a workload. We will show you what it costs here, and what it runs on, before you move it.

RevIssueHallsGPU
APod 01 energised, code hall only1384
BVision and knowledge halls in service31,152
CBiological design hall in service41,536