Assessment: Phoeniqs / Kvant Active Models for Conduct

3 min read

Assessment: Phoeniqs / Kvant Active Models for Conduct

This page assesses the models currently live on the Phoeniqs Model Service (Kvant Cloud) for use with Conduct. All models are served through an OpenAI-compatible API, so the deciding factors are capability fit (multimodal / vision, reasoning, tool use), context window, and throughput — not integration effort.

Bottom line: For top Conduct performance the primary model must be a vision / multimodal model. Text-only chat models are not sufficient — in an earlier client chat deployment, GPT-OSS was used and did not perform well. Best candidates: Qwen3-VL-235B (strongest vision-language model in the catalog) and Llama 4 Maverick (best for long-context multimodal workloads), with Gemma 4 31B as an efficient multimodal option.

Capability matrix

Legend — Conduct fit: Best candidate Secondary Not suitable as primary. Vision: ✅ supported ❌ not supported. “Relative to GPT-5.5” is an approximate overall capability level.

Model	Relative to GPT-5.5	Reasoning	Coding	General Chat	Long Context	Vision	Conduct fit	Overall
GPT OSS 120B	~85–90%	Excellent	Excellent	Excellent	Very Good (131k)	❌	Not suitable as primary	Best open model overall, but text-only; underperformed in Unique AI Chat.
DeepSeek V3.2	~85–90%	Excellent	Excellent	Excellent	Excellent (164k)	❌	Not suitable as primary	One of the strongest open models — text-only.
GLM 4.5 Air 110B	~80–85%	Very Good	Very Good	Excellent	Fair (8k)	❌	Not suitable as primary	Strong multilingual reasoning model; tiny 8k context.
Llama 4 Maverick	~75–80%	Good	Good	Very Good	Outstanding (1M)	✅	Best candidate	Best for long-context multimodal workloads.
Qwen3 VL 235B	~80–85%	Very Good	Good	Very Good	Very Good (65k)	✅	Best candidate	Strongest vision-language model in the catalog.
Gemma 4 31B	~75–80%	Good	Good	Good	Very Good (131k)	✅	Best candidate	Strong multimodal model with good efficiency.
Apertus 70B	~75–80%	Good	Good	Excellent	Good (56k)	❌	Not suitable as primary	Optimized for multilingual dialogue.
QwQ 32B	~80–85%	Excellent	Good	Moderate	Good (33k)	❌	Not suitable as primary	Specialized reasoning model.
Qwen3 8B	~60–65%	Good	Moderate	Good	Good (41k)	❌	Not suitable as primary	Excellent small reasoning model.
Gemma 3 12B IT	—	Good	Moderate	Good	Very Good (131k)	✅	Secondary	Solid multimodal, a step below Gemma 4 31B.
Llama 4 Scout 17B	—	Good	Good	Good	Good (62k)	✅	Secondary	Efficient multimodal fallback; smaller context than Maverick.
Granite Vision 3.2 2B	—	Basic	Weak	Basic	Fair (8k)	✅	Secondary	Compact VLM for narrow, high-throughput vision tasks.
Mistral 7B v0.3	—	Moderate	Moderate	Good	Good (32k)	❌	Not suitable as primary	Small multilingual chat model.
Granite 3.3 8B	—	Good	Moderate	Good	Good (32k)	❌	Not suitable as primary	Small instruction-following model.

Rows above the divider (GPT OSS 120B → Qwen3 8B) use the internal capability benchmark. Rows with “—” in the GPT-5.5 column are not yet benchmarked; their capability ratings are qualitative estimates.

Supporting stack (not primary Conduct models)

These pipeline components support a multimodal primary model rather than acting as the Conduct workhorse.

Model	Type	Role in Conduct
DeepSeek OCR	OCR	Document text extraction feeding the primary model.
MinerU 2.5	Vision-language (doc parsing)	High-accuracy document parsing / layout at the pre-processing stage.
BGE M3	Embedding	RAG retrieval embeddings.
Granite Embedding 278M	Embedding	Lightweight multilingual embeddings.
BGE Reranker v2 M3	Reranker	Relevance reranking in RAG.
Whisper Large v3	Speech-to-text	Transcription for audio inputs.

Recommendation

Standardise Conduct on a multimodal primary model. Use Qwen3-VL-235B where quality matters most, and Llama 4 Maverick when documents are long or image-heavy (1M-token context, up to 4 images per prompt). Gemma 4 31B is a good cost/latency-conscious default. Pair the primary model with the supporting stack — DeepSeek OCR / MinerU for document parsing, BGE M3 + BGE Reranker for RAG, Whisper for audio — rather than treating any single text-only chat model as the workhorse.

Caveats

Throughput figures on the source page are vendor benchmarks under controlled conditions and will vary in production. Phoeniqs reserves the right to change or deprecate models. Model licenses vary (Apache 2.0, MIT, Llama Community, Gemma, AGPL) and should be reviewed before production use. This assessment should be validated with a task-specific evaluation on representative Conduct workloads before committing.

Sources: Phoeniqs Model Service — Active Models; internal evaluation input (Slack, 2026-07-01).