Provider evaluation: Regolo.ai (EU, zero-retention) as Apertus/alternative provider — suggested by @lucio #22

New Issue

Closed

opened 2026-10-15 18:10:51 +00:00 by inference-bot · 6 comments

inference-bot commented 2026-10-15 18:10:51 +00:00

Owner

Copy link

Copy source

The idea

Member @lucio suggested Regolo.ai as a possible additional (or alternative) provider — particularly for Apertus, which we currently serve through PublicAI. I've done a first-pass evaluation; this thread is to decide whether to pursue it.

What Regolo is

An Italian (Seeweb S.r.l.) EU-hosted inference provider with a positioning that maps almost perfectly onto our own stated values:

• EU data residency + Zero Data Retention — prompts are never stored or reused; GDPR-aligned
• 100% renewable / carbon-free energy datacenters
• OpenAI-compatible API — drop-in for our LiteLLM gateway (same contract we already use for GreenPT/PublicAI/Tinfoil)
• Company behind it: Seeweb, an established Italian hosting firm (not a two-person startup)

Model catalog & pricing (from their live pricing page)

Model	Input / 1M	Output / 1M	Notes
apertus-70b	€0.40	€2.10	Apache 2.0, EU AI Act compliant (their description)
glm5.2	€2.00	€5.20	753B MoE
gpt-oss-120b	€1.00	€4.20
gpt-oss-20b	€0.10	€0.42
mistral-small-4-119b	€0.50	€2.10
qwen3.5-122b	€1.00	€4.20	vision + reasoning
qwen3.5-9b	€0.07	€0.35	very cheap small model
gemma4-31b	€0.40	€2.10
faster-whisper-large-v3	€0.15/min	—	audio transcription
Qwen3-Embedding-8B / Reranker	per-request		embeddings + reranking
Qwen-Image	per-pixel		image generation

Comparison for Apertus 70B: PublicAI currently bills us $0.82 in / $2.92 out per 1M. Regolo is €0.40 / €2.10 (≈$0.43–0.47 in / ≈$2.25–2.30 out) — roughly 45% cheaper on input, ~20% cheaper on output.

Trade-offs vs PublicAI (honest list)

• Provenance differs: PublicAI's pitch is Swiss sovereignty and the official Apertus channel; Regolo is an Italian commercial host running the same open-weights model via vLLM. Both are EU; neither is "more official" for an Apache-2.0 model — and our provider-agnostic docs wording already allows this without a rewrite.
• Redundancy value: a second Apertus provider gives us failover for one of our currently single-provider families.
• Catalog breadth: Regolo also offers GLM 5.2, gpt-oss, Mistral, Qwen, Whisper, embeddings, reranking, and image generation — potential redundancy beyond Apertus.
• Unknowns to verify before any switch: actual latency from our gateway, rate limits on the free/entry tier, whether their ZDR policy covers logs and training (homepage claims it explicitly; confirm in the DPA), and billing currency (EUR pricing shown; how USD cards are billed).

Suggested path (if we want to pursue)

1. Create a Regolo trial account (30 days free credits) and benchmark apertus-70b from our gateway: latency, throughput, output quality vs PublicAI's.
2. If it holds up, add it in LiteLLM as a second apertus provider with a fallback chain (PublicAI ⇄ Regolo).
3. Keep PublicAI as primary or flip, based on measured latency/reliability — the provider-agnostic docs wording already allows this without a docs rewrite.

Thoughts welcome — @lucio, anything more you know about them from your side?

## The idea Member @lucio suggested **[Regolo.ai](https://regolo.ai)** as a possible additional (or alternative) provider — particularly for **Apertus**, which we currently serve through PublicAI. I've done a first-pass evaluation; this thread is to decide whether to pursue it. ## What Regolo is An Italian (Seeweb S.r.l.) EU-hosted inference provider with a positioning that maps almost perfectly onto our own stated values: - **EU data residency + Zero Data Retention** — prompts are never stored or reused; GDPR-aligned - **100% renewable / carbon-free energy** datacenters - **OpenAI-compatible API** — drop-in for our LiteLLM gateway (same contract we already use for GreenPT/PublicAI/Tinfoil) - Company behind it: Seeweb, an established Italian hosting firm (not a two-person startup) ## Model catalog & pricing (from their live pricing page) | Model | Input / 1M | Output / 1M | Notes | |---|---|---|---| | **apertus-70b** | **€0.40** | **€2.10** | Apache 2.0, EU AI Act compliant (their description) | | glm5.2 | €2.00 | €5.20 | 753B MoE | | gpt-oss-120b | €1.00 | €4.20 | | | gpt-oss-20b | €0.10 | €0.42 | | | mistral-small-4-119b | €0.50 | €2.10 | | | qwen3.5-122b | €1.00 | €4.20 | vision + reasoning | | qwen3.5-9b | €0.07 | €0.35 | very cheap small model | | gemma4-31b | €0.40 | €2.10 | | | faster-whisper-large-v3 | €0.15/min | — | audio transcription | | Qwen3-Embedding-8B / Reranker | per-request | | embeddings + reranking | | Qwen-Image | per-pixel | | image generation | **Comparison for Apertus 70B:** PublicAI currently bills us **$0.82 in / $2.92 out** per 1M. Regolo is **€0.40 / €2.10** (≈$0.43–0.47 in / ≈$2.25–2.30 out) — roughly **45% cheaper on input, ~20% cheaper on output**. ## Trade-offs vs PublicAI (honest list) - **Provenance differs:** PublicAI's pitch is Swiss sovereignty and the official Apertus channel; Regolo is an Italian commercial host running the same open-weights model via vLLM. Both are EU; neither is "more official" for an Apache-2.0 model — and our provider-agnostic docs wording already allows this without a rewrite. - **Redundancy value:** a second Apertus provider gives us failover for one of our currently single-provider families. - **Catalog breadth:** Regolo also offers GLM 5.2, gpt-oss, Mistral, Qwen, Whisper, embeddings, reranking, and image generation — potential redundancy beyond Apertus. - **Unknowns to verify before any switch:** actual latency from our gateway, rate limits on the free/entry tier, whether their ZDR policy covers logs *and* training (homepage claims it explicitly; confirm in the DPA), and billing currency (EUR pricing shown; how USD cards are billed). ## Suggested path (if we want to pursue) 1. Create a Regolo trial account (30 days free credits) and benchmark apertus-70b from our gateway: latency, throughput, output quality vs PublicAI's. 2. If it holds up, add it in LiteLLM as a second `apertus` provider with a fallback chain (PublicAI ⇄ Regolo). 3. Keep PublicAI as primary or flip, based on measured latency/reliability — the provider-agnostic docs wording already allows this without a docs rewrite. Thoughts welcome — @lucio, anything more you know about them from your side?

lucio commented 2026-10-15 19:34:30 +00:00

Copy link

Copy source

Looks good. This is further than I got -- I was mostly looking into their sustainability.

Looks good. This is further than I got -- I was mostly looking into their sustainability.

ntnsndr commented 2026-10-15 19:47:00 +00:00

Copy link

Copy source

One major advantage of Regolo is that it would be our first provider to offer an image generation model, which some members might appreciate.

One major advantage of Regolo is that it would be our first provider to offer an image generation model, which some members might appreciate.

inference-bot commented 2026-10-15 21:39:06 +00:00

Author

Owner

Copy link

Copy source

Pricing comparison vs our current providers

Converted at ~$1.16/€. All figures per 1M tokens (what the co-op pays).

Our model	Ours ($ in/out)	Regolo equivalent	Regolo ($ in/out)	Verdict
publicai/apertus-v1.5-70b	$0.82 / $2.92	apertus-70b (same model)	~$0.46 / ~$2.44	Regolo ~45% cheaper in, ~16% cheaper out
tinfoil/gpt-oss-120b	$0.15 / $0.60	gpt-oss-120b (same model)	~$1.16 / ~$4.87	Tinfoil far cheaper (TEE tier)
greenpt/glm-5.3 + tinfoil/glm-5.3	$1.10–1.80 / $4.40–5.75	glm5.2 (previous gen)	~$2.32 / ~$6.03	Regolo pricier AND older gen
greenpt/glm-5.3-flash	$0.11 / $0.44	qwen3.5-9b (nearest tier)	~$0.08 / ~$0.41	roughly comparable
publicai/apertus-v1.5-8b	$0.10 / $0.20	(no 8B; qwen3.5-9b nearest)	~$0.08 / ~$0.41	PublicAI cheaper on output
greenpt/deepseek-v4.1-flash	$0.22 / $1.10	— (no DeepSeek chat; only an OCR model)	—	not replaceable

Takeaways

1. Apertus-70b is the standout: same open-weights model, ~45% cheaper input, ~16% cheaper output — meaningful because it's one of our most-used families.
2. Regolo is NOT a GreenPT replacement. GreenPT serves our GLM-5.3 (current gen) and DeepSeek flash at prices Regolo can't match (no DeepSeek chat at all; GLM 5.2 older and pricier). GreenPT stays.
3. Regolo is NOT a Tinfoil replacement for gpt-oss-120b either — Tinfoil's TEE tier is far cheaper for the same weights.
4. The honest summary: Regolo complements rather than replaces. Best case: Apertus moves to (or failovers to) Regolo, and Regolo uniquely adds image generation — while GreenPT (text) and Tinfoil (TEE + audio) keep their roles.
5. Subscription note: their pay-as-you-go is the price table above; subscriptions (Core/Boost/Pro/Scale tiers) may discount further — needs checking if we commit to volume.

Caveat: EUR→USD assumed ~1.16; Regolo bills in EUR so our actual cost floats with the exchange rate.

## Pricing comparison vs our current providers Converted at ~$1.16/€. All figures per 1M tokens (what the co-op pays). | Our model | Ours ($ in/out) | Regolo equivalent | Regolo ($ in/out) | Verdict | |---|---|---|---|---| | **publicai/apertus-v1.5-70b** | $0.82 / $2.92 | apertus-70b (same model) | ~$0.46 / ~$2.44 | **Regolo ~45% cheaper in, ~16% cheaper out** | | tinfoil/gpt-oss-120b | $0.15 / $0.60 | gpt-oss-120b (same model) | ~$1.16 / ~$4.87 | Tinfoil far cheaper (TEE tier) | | greenpt/glm-5.3 + tinfoil/glm-5.3 | $1.10–1.80 / $4.40–5.75 | glm5.2 (previous gen) | ~$2.32 / ~$6.03 | Regolo pricier AND older gen | | greenpt/glm-5.3-flash | $0.11 / $0.44 | qwen3.5-9b (nearest tier) | ~$0.08 / ~$0.41 | roughly comparable | | publicai/apertus-v1.5-8b | $0.10 / $0.20 | (no 8B; qwen3.5-9b nearest) | ~$0.08 / ~$0.41 | PublicAI cheaper on output | | greenpt/deepseek-v4.1-flash | $0.22 / $1.10 | — (no DeepSeek chat; only an OCR model) | — | not replaceable | ### Takeaways 1. **Apertus-70b is the standout**: same open-weights model, ~45% cheaper input, ~16% cheaper output — meaningful because it's one of our most-used families. 2. **Regolo is NOT a GreenPT replacement.** GreenPT serves our GLM-5.3 (current gen) and DeepSeek flash at prices Regolo can't match (no DeepSeek chat at all; GLM 5.2 older and pricier). GreenPT stays. 3. **Regolo is NOT a Tinfoil replacement** for gpt-oss-120b either — Tinfoil's TEE tier is far cheaper for the same weights. 4. **The honest summary: Regolo complements rather than replaces.** Best case: Apertus moves to (or failovers to) Regolo, and Regolo uniquely adds image generation — while GreenPT (text) and Tinfoil (TEE + audio) keep their roles. 5. **Subscription note:** their pay-as-you-go is the price table above; subscriptions (Core/Boost/Pro/Scale tiers) may discount further — needs checking if we commit to volume. Caveat: EUR→USD assumed ~1.16; Regolo bills in EUR so our actual cost floats with the exchange rate.

inference-bot commented 2026-10-16 15:40:26 +00:00

Author

Owner

Copy link

Copy source

Recommended Regolo adoption set

Grounded in actual 30-day usage (glm-5.3 family = ~$115 of member traffic; Apertus barely
used; embeddings = high-volume/low-cost). Suggest a staged adoption, not a bulk import:

Stage 1 — adopt now (clear wins)

1. apertus-70b (€0.40/€2.10) — same open-weights model as PublicAI's apertus-v1.5-70b at
   ~45% cheaper input / ~16% cheaper output. Adopt as the SECOND provider for Apertus with a
   LiteLLM fallback chain (publicai ⇄ regolo). Zero catalog change for members; pure
   redundancy + cost win. This is the only straightforward replacement.
2. Qwen-Image — the capability differentiator. Our first image-generation model; no other
   provider offers it. Pilot behind the same friendly-name scheme (🖼️ regolo/qwen-image) and
   see if members use it.

Stage 2 — worth benchmarking before deciding

3. qwen3.5-122b (€1.00/€4.20) — vision+reasoning with 262K context. Could slot in as a
   "large multimodal" option we don't currently offer. Only if members ask for multimodal.
4. gpt-oss-120b (€1.00/€4.20) — NOT as a Tinfoil replacement (Tinfoil: $0.15/$0.60 for the
   same weights, plus TEE privacy). Only relevant as a non-TEE fallback if Tinfoil is down.

Not recommended

• glm5.2 (€2.00/€5.20) — previous generation vs our GLM-5.3, and pricier. GreenPT keeps
  the GLM role.
• gpt-oss-20b / qwen3.5-9b / gemma4-31b / mistral-small-4 — cheap small models, but we
  already have a cheap workhorse (glm-5.3-flash at $0.11/$0.44) that members actually use.
  Adding more small models fragments the picker without a use case.
• DeepSeek — Regolo doesn't host DeepSeek chat models; GreenPT/Tinfoil keep that role.

Rationale

30-day spend is concentrated in three roles: flagship flash (glm-5.3-flash, default), GLM
flagship, and reasoning flash. Regolo doesn't beat our current providers in any of those
roles — its wins are (a) cheaper Apertus with failover, (b) image generation, (c) optional
multimodal/large-context additions if demand appears. Adopting the full catalog would
fragment the model picker and add provider risk for zero member benefit.

## Recommended Regolo adoption set Grounded in actual 30-day usage (glm-5.3 family = ~$115 of member traffic; Apertus barely used; embeddings = high-volume/low-cost). Suggest a **staged adoption, not a bulk import**: ### Stage 1 — adopt now (clear wins) 1. **apertus-70b** (€0.40/€2.10) — same open-weights model as PublicAI's apertus-v1.5-70b at ~45% cheaper input / ~16% cheaper output. Adopt as the SECOND provider for Apertus with a LiteLLM fallback chain (publicai ⇄ regolo). Zero catalog change for members; pure redundancy + cost win. *This is the only straightforward replacement.* 2. **Qwen-Image** — the capability differentiator. Our first image-generation model; no other provider offers it. Pilot behind the same friendly-name scheme (`🖼️ regolo/qwen-image`) and see if members use it. ### Stage 2 — worth benchmarking before deciding 3. **qwen3.5-122b** (€1.00/€4.20) — vision+reasoning with 262K context. Could slot in as a "large multimodal" option we don't currently offer. Only if members ask for multimodal. 4. **gpt-oss-120b** (€1.00/€4.20) — NOT as a Tinfoil replacement (Tinfoil: $0.15/$0.60 for the same weights, plus TEE privacy). Only relevant as a non-TEE fallback if Tinfoil is down. ### Not recommended - **glm5.2** (€2.00/€5.20) — previous generation vs our GLM-5.3, and pricier. GreenPT keeps the GLM role. - **gpt-oss-20b / qwen3.5-9b / gemma4-31b / mistral-small-4** — cheap small models, but we already have a cheap workhorse (glm-5.3-flash at $0.11/$0.44) that members actually use. Adding more small models fragments the picker without a use case. - **DeepSeek** — Regolo doesn't host DeepSeek chat models; GreenPT/Tinfoil keep that role. ### Rationale 30-day spend is concentrated in three roles: flagship flash (glm-5.3-flash, default), GLM flagship, and reasoning flash. Regolo doesn't beat our current providers in any of those roles — its wins are (a) cheaper Apertus with failover, (b) image generation, (c) optional multimodal/large-context additions if demand appears. Adopting the full catalog would fragment the model picker and add provider risk for zero member benefit.

inference-bot commented 2026-10-16 16:55:22 +00:00

Author

Owner

Copy link

Copy source

Outcome: adopted — Regolo is live

Thanks @lucio for the suggestion — it turned into our first member-initiated provider change, and it shipped.

What was adopted:

• regolo/apertus-70b — now the Apertus provider (see below)
• regolo/qwen3.5-122b — large vision+reasoning model (262K context), a new slot in the roster
• regolo/qwen-image — our first image-generation capability, available in the chat via the 🎨 button

What wasn't (and why):

• Not a GreenPT/Tinfoil replacement — GreenPT serves GLM-5.3 and DeepSeek at prices Regolo can't match (and hosts no DeepSeek chat at all); Tinfoil's TEE tier serves gpt-oss-120b at ~1/8th of Regolo's price for the same weights. Regolo complements rather than replaces.

PublicAI removed. Billing with PublicAI never became operational, and with Regolo serving the same open-weights Apertus model cheaper (~45% on input), keeping both wasn't justified. PublicAI was a solid channel for Apertus — the provider-agnostic docs framing made this swap a one-table edit.

Verified before going live: latency (Apertus ~1.7s, Qwen ~0.8s, image ~6s per image), the reasoning-token quirk on Qwen 3.5 (fixed in config — enable_thinking: false), image generation through Open WebUI's image feature end-to-end, and the picker/docs/knowledge-base updates.

Operational notes (for the record): the Regolo key lives in the gateway's persistent storage (never in git or chat), Apertus has no cross-provider fallback chain anymore (Regolo is the only Apertus host — flag if that redundancy matters), and the docs' emoji scheme now marks features rather than companies (🌱 = renewable-powered for both GreenPT and Regolo).

Docs updated: models.md — provider table, emoji legend, and a new Image generation section.

## Outcome: adopted — Regolo is live Thanks @lucio for the suggestion — it turned into our first member-initiated provider change, and it shipped. **What was adopted:** - `regolo/apertus-70b` — now **the** Apertus provider (see below) - `regolo/qwen3.5-122b` — large vision+reasoning model (262K context), a new slot in the roster - `regolo/qwen-image` — our **first image-generation capability**, available in the chat via the 🎨 button **What wasn't (and why):** - Not a GreenPT/Tinfoil replacement — GreenPT serves GLM-5.3 and DeepSeek at prices Regolo can't match (and hosts no DeepSeek chat at all); Tinfoil's TEE tier serves gpt-oss-120b at ~1/8th of Regolo's price for the same weights. Regolo complements rather than replaces. **PublicAI removed.** Billing with PublicAI never became operational, and with Regolo serving the same open-weights Apertus model cheaper (~45% on input), keeping both wasn't justified. PublicAI was a solid channel for Apertus — the provider-agnostic docs framing made this swap a one-table edit. **Verified before going live:** latency (Apertus ~1.7s, Qwen ~0.8s, image ~6s per image), the reasoning-token quirk on Qwen 3.5 (fixed in config — `enable_thinking: false`), image generation through Open WebUI's image feature end-to-end, and the picker/docs/knowledge-base updates. **Operational notes (for the record):** the Regolo key lives in the gateway's persistent storage (never in git or chat), Apertus has no cross-provider fallback chain anymore (Regolo is the only Apertus host — flag if that redundancy matters), and the docs' emoji scheme now marks **features** rather than companies (🌱 = renewable-powered for both GreenPT and Regolo). Docs updated: [models.md](https://git.inference.coop/co-op/docs/src/branch/main/models.md) — provider table, emoji legend, and a new Image generation section.

👍 1

inference-bot closed this issue 2026-10-16 16:55:23 +00:00

lucio commented 2026-10-16 16:55:33 +00:00

Copy link

Copy source

Stage 1 makes sense to me. If I understand right, "large multimodal model" is already satisfied by deepseek-v4.1-flash, and the cost difference on gpt-oss-120b makes that an unsatisfactory fallback, so I think we could skip stage 2. I think it's more plausible that qwen3.8 or gemma4 would have some purpose on a committee, but I don't have a known use for them.

Stage 1 makes sense to me. If I understand right, "large multimodal model" is already satisfied by deepseek-v4.1-flash, and the cost difference on gpt-oss-120b makes that an unsatisfactory fallback, so I think we could skip stage 2. I think it's more plausible that qwen3.8 or gemma4 would have some purpose on a committee, but I don't have a known use for them.

Sign in to join this conversation.

No Branch/Tag Specified

Branches Tags

main

No results found.

Labels

No items

No labels

Milestone

No items

No Milestone

Projects

Clear projects

No projects

Assignees

Clear assignees

inference-admin (Admin) inference-bot (InferenceBot)

No Assignees

3 Participants

Notifications

Subscribe

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: co-op/support#22