Free sweep, 2026-11-03: what changed, and what could add capacity #369

Closed
opened 2026-11-03 11:39:22 -05:00 by cmoriarty · 10 comments
Owner

The sweep of free plans, 2026-11-03. Some numbers come from community lists, marked below. New models need a recall probe before they join a chain.

The problem is capacity. Run 170 asked for about 410 strong requests an hour, averaging 54K input tokens (469 requests and 25.2M input tokens in 64.5 minutes, before most of its steps ran).

Strong-tier model on production What it can give What happened on run 170
Kilo beam-free About 100 an hour (measured, #367) Out after 12 minutes
OVH Qwen 397B 120 an hour (2 a minute) The only model left after minute 45
Zen About 184, then blocked for 24 h or more Blocked at minute 45
Cloudflare Gemma About 27 requests a day (1M tokens) Spent in one step
Gemini Flash-Lite 500 a day Held for vision and judge steps, by design

So the steady supply is about 220 an hour against 410 wanted, and a second run the same day gets less.

Changed since our catalog (2026-10-17)

  • Kilo: refuses at about 100 an hour; docs still say 200. New free models, all sharing that limit.
  • Zen: block came at 184, not 200. longcat-2.5-preview-free is gone. New: beam-free, mimo-v2.6-flash-free, flint-40b-free, pulse-alpha-free, two Nemotrons. Unknown whether the block is per model.
  • OVH: now also serves Qwen3.8-27B and Qwen3.6-27B (262K). Worth testing whether its 2 a minute is per model.
  • Inception: grant 100M since 10-19, limits 10x higher, new mercury-3 (1M). Our probe failed on 10-17 at capacity, so it's off. Key already on production.
  • Z.ai: glm-4.7-flash also failed its probe as overloaded, not on recall. Retest.
  • OpenRouter: limits unchanged. Live id is dots-studio/dots-3-note-preview:free; ours lacks the prefix, so check it. inkling free models end 11-04.
  • Gemini: limits are no longer published, only shown in AI Studio. gemini-3.6-flash is new on the free tier.
  • Alibaba: new activations get 90 days; free quota removed for many models on 10-29. Off, and stays off.
  • Unchanged: Cloudflare 10,000 neurons, Cohere 1,000 a month, Groq 8K tokens a minute (still under one request), NVIDIA off by its terms.

Worth adding

Provider Free limit Needs Catch
Hetzner Qwen3.8-27B (262K, images) 10 requests a minute, 4M input tokens a minute Account; one report says a card on file "Experimental", not for production, may end with notice. Content not stored
Mistral Experiment plan 1 a second, 500K tokens a minute per model; monthly cap only from third parties Phone Prompts may train. Console refused keys on 10-17; retry
Nous Portal: Beam, Ornith 1.5 Daily quotas, unpublished Account; one report of a $0 card step Test it
Kilo key Docs say same limit; the refusal says sign up Account Test it (#367)

Hetzner alone (600 an hour) covers the demand, and it's the model production's main profile uses, so a run could stay on one model.

Later, or unknown: Vercel AI Gateway (three $0 models, limits unpublished), Amazon Nova (free tier, numbers not found), Requesty (200 a day).

Not worth it: AIHubMix (100 a day, shared), IteraCompute (131K window, 8K output), Dahl (one-time, data use unstated, unknown GPU hosts), ModelScope (real-name check), Cerebras (card), and the many lists now marked no longer free (GitHub, Together, SambaNova, Hugging Face, DeepInfra, DeepSeek, Chutes).

Avoid: LLM7 and similar routers offering paid frontier models free, with little or no sign-up. Who pays, and who reads the prompts, can't be checked.

Done when

  • The catalog matches this list, each new model passes its probe, and a test covers it.
  • A full Free run has enough strong capacity to reach its Visual check.

Sources: Hetzner, Kilo, Zen, OpenRouter, Gemini, Cloudflare, Cohere, Groq, Alibaba, Inception, Vercel, models.dev, community: free-llm.com.

The sweep of free plans, 2026-11-03. Some numbers come from community lists, marked below. New models need a recall probe before they join a chain. **The problem is capacity.** Run 170 asked for about 410 strong requests an hour, averaging 54K input tokens (469 requests and 25.2M input tokens in 64.5 minutes, before most of its steps ran). | Strong-tier model on production | What it can give | What happened on run 170 | |---|---|---| | Kilo `beam-free` | About 100 an hour (measured, #367) | Out after 12 minutes | | OVH Qwen 397B | 120 an hour (2 a minute) | The only model left after minute 45 | | Zen | About 184, then blocked for 24 h or more | Blocked at minute 45 | | Cloudflare Gemma | About 27 requests a day (1M tokens) | Spent in one step | | Gemini Flash-Lite | 500 a day | Held for vision and judge steps, by design | So the steady supply is about 220 an hour against 410 wanted, and a second run the same day gets less. **Changed since our catalog (2026-10-17)** - **Kilo:** refuses at about 100 an hour; docs still say 200. New free models, all sharing that limit. - **Zen:** block came at 184, not 200. `longcat-2.5-preview-free` is gone. New: `beam-free`, `mimo-v2.6-flash-free`, `flint-40b-free`, `pulse-alpha-free`, two Nemotrons. Unknown whether the block is per model. - **OVH:** now also serves `Qwen3.8-27B` and `Qwen3.6-27B` (262K). Worth testing whether its 2 a minute is per model. - **Inception:** grant 100M since 10-19, limits 10x higher, new `mercury-3` (1M). Our probe failed on 10-17 at capacity, so it's off. Key already on production. - **Z.ai:** `glm-4.7-flash` also failed its probe as overloaded, not on recall. Retest. - **OpenRouter:** limits unchanged. Live id is `dots-studio/dots-3-note-preview:free`; ours lacks the prefix, so check it. `inkling` free models end 11-04. - **Gemini:** limits are no longer published, only shown in AI Studio. `gemini-3.6-flash` is new on the free tier. - **Alibaba:** new activations get 90 days; free quota removed for many models on 10-29. Off, and stays off. - **Unchanged:** Cloudflare 10,000 neurons, Cohere 1,000 a month, Groq 8K tokens a minute (still under one request), NVIDIA off by its terms. **Worth adding** | Provider | Free limit | Needs | Catch | |---|---|---|---| | **Hetzner** `Qwen3.8-27B` (262K, images) | 10 requests a minute, 4M input tokens a minute | Account; one report says a card on file | "Experimental", not for production, may end with notice. Content not stored | | **Mistral** Experiment plan | 1 a second, 500K tokens a minute per model; monthly cap only from third parties | Phone | Prompts may train. Console refused keys on 10-17; retry | | **Nous Portal**: Beam, Ornith 1.5 | Daily quotas, unpublished | Account; one report of a $0 card step | Test it | | **Kilo key** | Docs say same limit; the refusal says sign up | Account | Test it (#367) | Hetzner alone (600 an hour) covers the demand, and it's the model production's `main` profile uses, so a run could stay on one model. **Later, or unknown:** Vercel AI Gateway (three $0 models, limits unpublished), Amazon Nova (free tier, numbers not found), Requesty (200 a day). **Not worth it:** AIHubMix (100 a day, shared), IteraCompute (131K window, 8K output), Dahl (one-time, data use unstated, unknown GPU hosts), ModelScope (real-name check), Cerebras (card), and the many lists now marked no longer free (GitHub, Together, SambaNova, Hugging Face, DeepInfra, DeepSeek, Chutes). **Avoid:** LLM7 and similar routers offering paid frontier models free, with little or no sign-up. Who pays, and who reads the prompts, can't be checked. **Done when** - The catalog matches this list, each new model passes its probe, and a test covers it. - A full Free run has enough strong capacity to reach its Visual check. Sources: [Hetzner](https://docs.hetzner.com/general/company-and-policy/experiments/inference/), [Kilo](https://kilo.ai/docs/gateway/usage-and-billing), [Zen](https://opencode.ai/docs/zen/), [OpenRouter](https://openrouter.ai/docs/api/reference/limits), [Gemini](https://aistudio.google.com/docs/rate-limits), [Cloudflare](https://developers.cloudflare.com/workers-ai/platform/pricing/), [Cohere](https://docs.cohere.com/docs/rate-limits), [Groq](https://console.groq.com/docs/rate-limits), [Alibaba](https://www.alibabacloud.com/help/en/model-studio/new-free-quota), [Inception](https://docs.inceptionlabs.ai/get-started/rate-limits), [Vercel](https://vercel.com/docs/ai-gateway/rate-limits), [models.dev](https://models.dev/api.json), community: [free-llm.com](https://free-llm.com/).
Author
Owner

Corrections after checking the sign-up pages

  • Hetzner: its fraud prevention FAQ says a new account may have to verify with a government ID or an advance card payment. The amount isn't stated, and unused credit comes back only when the account is closed. So the API is free, but getting in may not be. Tokens are made at experiments.hetzner.com/inference.
  • Mistral: the official quickstart now says API access is on by default with no card, and doesn't mention a phone. An older help article does. Real limits show only on the admin Limits page. The monthly cap is still third-party only.
  • Nous: Ornith isn't free. The live list's $0 models are Beam (262K), Step 3.7 Flash and Solar Mini 4. The portal says API keys use credits, but a June report got a 200 from a :free model with one. So it may or may not work; test it.
  • Kilo: its docs say the free limit is per IP for signed-in accounts too, so a key may not help.
  • New keys: keyenv.py only knows KILO_API_KEY so far. HETZNER_API_KEY, MISTRAL_API_KEY and NOUS_API_KEY need adding there, and all of them in deploy/docker-compose.yml, before production can use them.
**Corrections after checking the sign-up pages** - **Hetzner:** its [fraud prevention FAQ](https://docs.hetzner.com/general/security-and-identify/fraud-prevention-faq/) says a new account may have to verify with a government ID or an advance card payment. The amount isn't stated, and unused credit comes back only when the account is closed. So the API is free, but getting in may not be. Tokens are made at `experiments.hetzner.com/inference`. - **Mistral:** the [official quickstart](https://docs.mistral.ai/getting-started/quickstarts/studio/activate-and-generate-api-key/) now says API access is on by default with no card, and doesn't mention a phone. An older help article does. Real limits show only on the admin Limits page. The monthly cap is still third-party only. - **Nous:** Ornith isn't free. The live list's $0 models are Beam (262K), Step 3.7 Flash and Solar Mini 4. The portal says API keys use credits, but a [June report](https://github.com/NousResearch/hermes-agent/issues/47950) got a 200 from a `:free` model with one. So it may or may not work; test it. - **Kilo:** its [docs](https://kilo.ai/docs/gateway/usage-and-billing) say the free limit is per IP for signed-in accounts too, so a key may not help. - **New keys:** `keyenv.py` only knows `KILO_API_KEY` so far. `HETZNER_API_KEY`, `MISTRAL_API_KEY` and `NOUS_API_KEY` need adding there, and all of them in `deploy/docker-compose.yml`, before production can use them.
Author
Owner

Hetzner's token works (2026-11-03). One tiny request to Qwen3.8-27B returned 200, and it answered as Qwen/Qwen3.8-27B. Its list shows Qwen/Qwen3.6-35B-A3B, Qwen/Qwen3.6-35B-A3B-FP8, Qwen/Qwen3.8-27B and Qwen3.8-27B, so the ids differ from the docs: check which to use.

The account needed a payment method (PayPal was accepted, and its payments are manual) and a government ID. Limit headers: none returned, so the 10 a minute is still only what the docs say. Recall and tool probes are still to do.

**Hetzner's token works (2026-11-03).** One tiny request to `Qwen3.8-27B` returned 200, and it answered as `Qwen/Qwen3.8-27B`. Its list shows `Qwen/Qwen3.6-35B-A3B`, `Qwen/Qwen3.6-35B-A3B-FP8`, `Qwen/Qwen3.8-27B` and `Qwen3.8-27B`, so the ids differ from the docs: check which to use. The account needed a payment method (PayPal was accepted, and its payments are manual) and a government ID. Limit headers: none returned, so the 10 a minute is still only what the docs say. Recall and tool probes are still to do.
Author
Owner

Mistral is out (2026-11-03). On our account, the API keys page still says "Your organization is on the Free plan. Upgrade to activate your API keys and access the full API", with New key disabled. Mistral's quickstart says the opposite, and looks out of date: its API keys page says you cannot create a Free mode key. Two others report the same since October (one, two). Community lists still show about 1B tokens a month, which seems to reflect only older accounts with existing keys.

It's paid only now, and not worth it: a run like run 170 would cost about $13 in input tokens alone at Mistral Large's listed price.

So Hetzner is the only large new source, and it's experimental. That makes a second one, such as Nous's free Beam, worth more.

**Mistral is out (2026-11-03).** On our account, the API keys page still says "Your organization is on the Free plan. Upgrade to activate your API keys and access the full API", with New key disabled. Mistral's quickstart says the opposite, and looks out of date: [its API keys page](https://docs.mistral.ai/admin/security-access/api-keys) says you cannot create a Free mode key. Two others report the same since October ([one](https://github.com/mnfst/awesome-free-llm-apis/issues/213), [two](https://github.com/cheaper-inference/OmniRoute/issues/16279)). Community lists still show about 1B tokens a month, which seems to reflect only older accounts with existing keys. It's paid only now, and not worth it: a run like run 170 would cost about $13 in input tokens alone at Mistral Large's listed price. So Hetzner is the only large new source, and it's experimental. That makes a second one, such as Nous's free Beam, worth more.
Author
Owner

Nous is out, and Kilo's key works (2026-11-03).

  • Nous: the portal's API keys page says "You need available funds to create or use API keys", so its Free plan can't be used from Braid. Its free models look meant for its own Hermes app. Buying credit might unlock them, but the daily quotas are unpublished, so that's not worth paying to find out.
  • Kilo: a signed-in key got a 200 from beam-free. Whether it lifts the 100 an hour is still untested, and its docs say the limit stays per IP (#367).

Where that leaves new capacity: Hetzner is the only large new source. Its backups are the no-account retests: OVH's other Qwen models, Inception and Z.ai. Vercel and Amazon Nova stay in reserve if those fall short.

**Nous is out, and Kilo's key works (2026-11-03).** - **Nous:** the portal's API keys page says "You need available funds to create or use API keys", so its Free plan can't be used from Braid. Its free models look meant for its own Hermes app. Buying credit might unlock them, but the daily quotas are unpublished, so that's not worth paying to find out. - **Kilo:** a signed-in key got a 200 from `beam-free`. Whether it lifts the 100 an hour is still untested, and its docs say the limit stays per IP (#367). **Where that leaves new capacity:** Hetzner is the only large new source. Its backups are the no-account retests: OVH's other Qwen models, Inception and Z.ai. Vercel and Amazon Nova stay in reserve if those fall short.
Author
Owner

Nous confirmed out (2026-11-03). The $0 checkout works and puts the account on the Free plan, but API keys still can't be created. The plan page explains it: "API keys always use your credits (pay-as-you-go), not included usage", and the Free plan has $0 monthly credits. The plan is for Nous's own Hermes Agent app, signed in through OAuth. Pointing Braid at that sign-in would be using one product's access for another, so it's not an option.

**Nous confirmed out (2026-11-03).** The $0 checkout works and puts the account on the Free plan, but API keys still can't be created. The plan page explains it: "API keys always use your credits (pay-as-you-go), not included usage", and the Free plan has $0 monthly credits. The plan is for Nous's own Hermes Agent app, signed in through OAuth. Pointing Braid at that sign-in would be using one product's access for another, so it's not an option.
Author
Owner

Gemini's real free limits, read from AI Studio (2026-11-03). They match the catalog's third-party numbers, so those can be marked as read from the source.

Model Requests a minute Tokens a minute Requests a day
gemini-3.5-flash-lite 15 250K 500
gemini-3.1-flash-lite 15 250K 500
gemini-3.8-flash 5 250K 20
gemini-3.6-flash 5 250K 20
Gemma 4 26B and 31B 30 16K 14.4K
  • Catalog gaps: gemini-3.1-flash-lite and gemini-3.8-flash have no token limit recorded; add 250K.
  • Gemma on Gemini's API: no use, for the same reason as Groq. 16K a minute can't hold one Braid request (54K on average).
  • gemini-3.6-flash: only another 20 a day, so it could help the judge tier at most.
  • At these limits Flash-Lite would manage about 4.6 average requests a minute on tokens, not 15.
**Gemini's real free limits, read from AI Studio (2026-11-03).** They match the catalog's third-party numbers, so those can be marked as read from the source. | Model | Requests a minute | Tokens a minute | Requests a day | |---|---|---|---| | `gemini-3.5-flash-lite` | 15 | 250K | 500 | | `gemini-3.1-flash-lite` | 15 | 250K | 500 | | `gemini-3.8-flash` | 5 | 250K | 20 | | `gemini-3.6-flash` | 5 | 250K | 20 | | Gemma 4 26B and 31B | 30 | **16K** | 14.4K | - **Catalog gaps:** `gemini-3.1-flash-lite` and `gemini-3.8-flash` have no token limit recorded; add 250K. - **Gemma on Gemini's API:** no use, for the same reason as Groq. 16K a minute can't hold one Braid request (54K on average). - **`gemini-3.6-flash`:** only another 20 a day, so it could help the judge tier at most. - At these limits Flash-Lite would manage about 4.6 average requests a minute on tokens, not 15.
Author
Owner

Retests with tiny requests (2026-11-03)

  • OVH: five requests to four models within about 30 s drew no refusal, so its 2 a minute is per model. Qwen3.8-27B and Qwen3.6-27B add capacity with no key. It can be slow: one request timed out at 120 s, one took 25 s.
  • Inception: mercury-3 and mercury-2.5 answered in 0.5 s. The 10-17 probe failure is out of date, so probe again.
  • Z.ai: glm-4.7-flash still answers "overloaded" (1305). It stays out.
  • Zen: deepseek-v4.1-flash-free, beam-free and mimo-v2.6-flash-free all answered FreeUsageLimitError. The block covers every model, so its new free models add no capacity.
  • Shared IP: the laptop got Zen's block too, so it shares production's IP. Tests from here count against production's per-IP limits.
  • OpenRouter: dots-3-note-preview:free still answers, as an alias of the prefixed id. No fix needed.
  • Kilo's key: 299 requests, about three times the no-key count (#367).

The plan is in the OpenSpec change free-capacity-hetzner.

**Retests with tiny requests (2026-11-03)** - **OVH:** five requests to four models within about 30 s drew no refusal, so its 2 a minute is per model. `Qwen3.8-27B` and `Qwen3.6-27B` add capacity with no key. It can be slow: one request timed out at 120 s, one took 25 s. - **Inception:** `mercury-3` and `mercury-2.5` answered in 0.5 s. The 10-17 probe failure is out of date, so probe again. - **Z.ai:** `glm-4.7-flash` still answers "overloaded" (1305). It stays out. - **Zen:** `deepseek-v4.1-flash-free`, `beam-free` and `mimo-v2.6-flash-free` all answered `FreeUsageLimitError`. The block covers every model, so its new free models add no capacity. - **Shared IP:** the laptop got Zen's block too, so it shares production's IP. Tests from here count against production's per-IP limits. - **OpenRouter:** `dots-3-note-preview:free` still answers, as an alias of the prefixed id. No fix needed. - **Kilo's key:** 299 requests, about three times the no-key count (#367). The plan is in the OpenSpec change `free-capacity-hetzner`.
Author
Owner

Measured: Hetzner is steady but slow (2026-11-03). I replayed run 170's spec-test step with Hetzner at the head of the chain, and stopped it at 31 minutes with no file written yet.

Model Requests Speed Time per request
Hetzner Qwen3.8-27B (this replay) 14 20 tok/s 86 s
Kilo beam-free (production, last 6 days) 367 100 tok/s 4.4 s
OVH Qwen3.5-397B (production, last 6 days) 155 8 tok/s 44 s

Hetzner refused nothing. Two of its requests thought for about 8,000 tokens, 6.4 and 7.5 minutes each.

Its 10 a minute isn't the limit that counts: at 86 s a request, one session gets about 40 requests an hour.

So the order changes. Kilo leads the strong and reader chains while its count lasts, and Hetzner carries a step when Kilo refuses, ahead of OVH. A private repository can't use Kilo, so Hetzner leads there.

**Measured: Hetzner is steady but slow (2026-11-03).** I replayed run 170's spec-test step with Hetzner at the head of the chain, and stopped it at 31 minutes with no file written yet. | Model | Requests | Speed | Time per request | |---|---|---|---| | Hetzner Qwen3.8-27B (this replay) | 14 | 20 tok/s | 86 s | | Kilo beam-free (production, last 6 days) | 367 | 100 tok/s | 4.4 s | | OVH Qwen3.5-397B (production, last 6 days) | 155 | 8 tok/s | 44 s | Hetzner refused nothing. Two of its requests thought for about 8,000 tokens, 6.4 and 7.5 minutes each. Its 10 a minute isn't the limit that counts: at 86 s a request, one session gets about 40 requests an hour. **So the order changes.** Kilo leads the strong and reader chains while its count lasts, and Hetzner carries a step when Kilo refuses, ahead of OVH. A private repository can't use Kilo, so Hetzner leads there.
Author
Owner

Shipped in PR 371. Left open until a full Free run shows it holds, which is the plan's next step.

In the chains now:

  • Hetzner's Qwen3.8-27B, behind Kilo in strong and reader, and leading for a private repository. Its key is set on production.
  • OVH's Qwen3.8-27B and Qwen3.6-27B: its limit is per model, so each adds capacity with no key.
  • Inception's mercury-3.

In the catalog: Gemini's limits as AI Studio shows them, Zen's block at about 180 requests, and Zen's longcat-2.5-preview-free ended.

Same step, same capsule, two chain orders:

Leading the chain Result
Kilo beam-free Succeeded in 1.6 min
Hetzner Qwen3.8-27B No file written after 31 min, stopped

Left out, or not worth it: Mistral and Nous need payment for API keys, whatever their pages suggest. Gemma on Gemini's API has too few tokens a minute. Z.ai is still overloaded.

Owner decisions, not urgent: OpenRouter's one-time $10 for 1,000 requests a day, and NVIDIA's opt-in, whose terms don't allow auto-approve to stay on.

**Shipped in [PR 371](https://forgejo.underthere.xyz/cmoriarty/braid/pulls/371).** Left open until a full Free run shows it holds, which is the plan's next step. **In the chains now:** - Hetzner's Qwen3.8-27B, behind Kilo in strong and reader, and leading for a private repository. Its key is set on production. - OVH's Qwen3.8-27B and Qwen3.6-27B: its limit is per model, so each adds capacity with no key. - Inception's mercury-3. **In the catalog:** Gemini's limits as AI Studio shows them, Zen's block at about 180 requests, and Zen's longcat-2.5-preview-free ended. **Same step, same capsule, two chain orders:** | Leading the chain | Result | |---|---| | Kilo beam-free | Succeeded in 1.6 min | | Hetzner Qwen3.8-27B | No file written after 31 min, stopped | **Left out, or not worth it:** Mistral and Nous need payment for API keys, whatever their pages suggest. Gemma on Gemini's API has too few tokens a minute. Z.ai is still overloaded. **Owner decisions, not urgent:** OpenRouter's one-time $10 for 1,000 requests a day, and NVIDIA's opt-in, whose terms don't allow auto-approve to stay on.
Author
Owner

Closing: run 171 went end to end on the new chains, through the Visual check to a merged PR, in 104.7 minutes with nobody needed. Run 170 had stopped at its first spec-test step after 64.5.

  • Kilo first paid off: 260 requests, 3 s each on average, and the build took 5.7 minutes against 36.5.
  • Hetzner carried the judge steps: 141 requests, 22 s on average, with some thinking requests of 4 to 7 minutes (#366).
  • None refused, none parked.

Kilo didn't reach its count on this run, so running out rests on the sim free-kilo-refuses-early.

Left with the owner, not urgent: OpenRouter's $10 one-time for 1,000 a day, and NVIDIA's opt-in. Details of the run are on #365.

**Closing: run 171 went end to end on the new chains**, through the Visual check to a merged PR, in 104.7 minutes with nobody needed. Run 170 had stopped at its first spec-test step after 64.5. - Kilo first paid off: 260 requests, 3 s each on average, and the build took 5.7 minutes against 36.5. - Hetzner carried the judge steps: 141 requests, 22 s on average, with some thinking requests of 4 to 7 minutes (#366). - None refused, none parked. Kilo didn't reach its count on this run, so running out rests on the sim `free-kilo-refuses-early`. Left with the owner, not urgent: OpenRouter's $10 one-time for 1,000 a day, and NVIDIA's opt-in. Details of the run are on #365.
Sign in to join this conversation.
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid#369
No description provided.