Recipes_

Gpodz-curated LoRA recipes across Qwen3.5 (8 variants, Image-Text-to-Text), Gemma 4 (4 variants, Any-to-Any / Image-Text-to-Text), Gemma 3 No-Think (1 variant, latency-optimized), and DeepSeek V4 Flash (2 reserved-booking variants, text-only). Pick the base that fits your VRAM budget and archetype; Gpodz handles adapter composition, eval gating, and serving warm-up.

15 pinned recipes · indicative pricing from docs/21 §2 · billing starts only when your GPU lane passes readiness.

Available now — mvp

Qwen Starter 2B

Qwen/Qwen3.5-2B · 2B · dense

mvp

Modality: Image-Text-to-Text

Min lane: Shared · shared-12gb

Training (indicative): $5 setup + $0.45/GPU-hr

Serving (indicative): $0.30/hr (Quick)

Compact 2B-class Qwen recipe for cheap dev iteration, agent prototyping, and the Toolkit-Chat dogfood path.

Qwen Starter 4B

Qwen/Qwen3.5-4B · 4B · dense

mvp

Modality: Image-Text-to-Text

Min lane: Shared · shared-12gb

Training (indicative): $9 setup + $0.90/GPU-hr

Serving (indicative): $0.60/hr (Quick)

Curated Qwen3.5 4B recipe for first adapters, support bots, and simple agents on Shared or 23GB Isolated GPU capacity.

Qwen Starter 9B

Qwen/Qwen3.5-9B · 9B · dense

mvp

Modality: Image-Text-to-Text

Min lane: Shared · shared-16gb

Training (indicative): $9 setup + $0.90/GPU-hr

Serving (indicative): $1.05/hr (Quick)

Curated Qwen3.5 9B recipe for stronger first adapters, support bots, and small agents on 16GB Shared or 23GB Isolated capacity.

Qwen Pro 27B

Qwen/Qwen3.5-27B · 27B · dense

mvp

Modality: Image-Text-to-Text

Min lane: Isolated · isolated-45gb

Training (indicative): $29 setup + $2.40/GPU-hr

Serving (indicative): $1.80/hr (Quick)

Curated Qwen3.5 27B recipe for domain workflows and serious agents on 45GB Isolated capacity.

Qwen Max 35B

Qwen/Qwen3.5-35B-A3B · 35B-A3B · MoE (3B active)

mvp

Modality: Image-Text-to-Text

Min lane: Isolated · isolated-45gb

Training (indicative): $49 setup + $3.60/GPU-hr

Serving (indicative): $2.40/hr (Quick)

Curated Qwen3.5 35B-A3B MoE recipe for production agents and codebase, workflow, and domain adapters on 45GB Isolated capacity.

Qwen Mega 122B

Qwen/Qwen3.5-122B-A10B · 122B-A10B · MoE (10B active)

mvp

Modality: Image-Text-to-Text

Min lane: Dedicated · dedicated-80gb

Serving (indicative): Dedicated GPU rate — contact for pricing

Qwen3.5 MoE mega model at 122B total (10B active) for heavy inference on dedicated Hopper/Blackwell hardware.

Gemma Lite E4B

google/gemma-4-E4B-it · ~4B · efficient/edge — Any-to-Any

mvp

Modality: Any-to-Any (text · image · audio · video)

Min lane: Shared · shared-12gb

Training (indicative): $9 setup + $0.90/GPU-hr

Serving (indicative): $0.60/hr (Quick)

Curated Gemma 4 E4B recipe for compact assistants and lower-cost text experiments on Shared or 23GB Isolated GPU capacity.

Gemma Edge E2B

google/gemma-4-E2B · ~2B · efficient/edge — multimodal

mvp

Modality: Image-Text-to-Text

Min lane: Shared · shared-12gb

Serving (indicative): Shared GPU rate — contact for pricing

Gemma 4 E2B — smallest Gemma 4 variant for latency-sensitive T2/T3 lanes.

Gemma No-Think 27B

google/gemma-3-27b-it · 27B · dense — Gemma 3 (chain-of-thought disabled)

mvp

Modality: Image-Text-to-Text

Min lane: Isolated · isolated-45gb

Serving (indicative): Isolated GPU rate — contact for pricing

Gemma 3 27B IT — chain-of-thought disabled for latency-sensitive inference paths.

Beta — validation in progress

Beta recipes are available but carry a beta banner in the training wizard. Multimodal and long-context probes are still completing.

Gemma Vision 26B

google/gemma-4-26B-A4B-it · 26B-A4B · MoE (4B active) — Image-Text-to-Text

beta

Modality: Image-Text-to-Text

Min lane: Isolated · isolated-45gb

Training (indicative): $49 setup + $3.60/GPU-hr

Serving (indicative): $2.40/hr (Quick)

Curated Gemma 4 26B-A4B recipe for multimodal and domain adapters on 45GB Isolated capacity. Multimodal validation in progress.

Gemma Max 31B

google/gemma-4-31B-it · 31B · dense — Image-Text-to-Text

beta

Modality: Image-Text-to-Text

Min lane: Isolated · isolated-45gb

Training (indicative): $79 setup + $4.80/GPU-hr

Serving (indicative): $2.40/hr (Quick)

Curated Gemma 4 31B recipe for higher-quality text and vision adapters on 45GB Isolated capacity (90GB lane added after validation).

Reserved booking required

These recipes are not self-serve. They require a pre-arranged reservation on dedicated B200 or H200 capacity. Contact Gpodz to schedule. Idle-billing applies on hot-pool reservations.

DeepSeek Flash Reserved

deepseek-ai/DeepSeek-V4-Flash · Flash · dense (FP8)

reserved booking required

Modality: Text-only

Min lane: Dedicated · dedicated-180gb

Reserved booking required — $7.20/hr (indicative). Contact Gpodz to schedule a dedicated block.

Reserved DeepSeek V4 Flash serving block for long-context reasoning and code or document analysis on a dedicated B200 or H200. 4-hour minimum.

DeepSeek Flash Hot

deepseek-ai/DeepSeek-V4-Flash · Flash · dense (FP8)

reserved booking required

Modality: Text-only

Min lane: Dedicated · dedicated-180gb

Reserved booking required — daily rate, contact Gpodz. Base model is billable while idle (LEGAL-8).

DeepSeek V4 Flash hot pool for low-latency reserved serving on a dedicated B200 or H200. Base model stays resident and is billable while idle.

Internal / pipeline use

These recipes carry launch_status: manual_review and are NOT available to tenant principals. They require the internal:pipeline scope on an API key (CLAUDE.md §8). Shown here for operator visibility only.

Qwen Frontier 397B

manual_review

Qwen/Qwen3.5-397B-A17B · 397B-A17B · MoE (17B active) — multi-GPU tensor-parallel

Modality: Image-Text-to-Text

Requires internal:pipeline scope. Qwen3.5-397B-A17B MoE frontier model — reserved for multi-GPU tensor-parallel serving. Phase 1 does not ship this publicly.

Qwen Starter Pipeline (0.8B)

manual_review

Qwen/Qwen3.5-0.8B · 0.8B · dense — pipeline test only

Modality: Image-Text-to-Text

Requires internal:pipeline scope. Pipeline-test recipe for end-to-end platform validation on the smallest Qwen-line model. Internal only. Never invoiced.

Failed readiness ⇒ no charge. Billing starts only when your GPU lane passes readiness. See trust page for the Gate-4 billing proof.

Indicative pricing shown on each card. Final rates are the authoritative billing-engine values per docs/21 §2 — not this YAML hint. All 15 recipes are pinned to real HuggingFace revision SHAs verified 2026-05-14.

Subscribe to recipe updates_