# Model updates checked September 25, 2026

**Unreleased source catalog.** These additions are not in the v0.2.2 download or
the existing v0.3.0-rc.1 draft. All four routes passed a small live text/tool-call
check on September 22, 2026: read a harmless file and return its marker word.
GLM required another run after initial connection failures; those failed attempts
did not report usage, so their cost is unknown. This verifies that
small workflow, not general reliability, image support or comparative task quality.

| New route | Why consider it | Current integration |
| --- | --- | --- |
| `mimo-v2.6-flash` | Lower token price in the new MiMo family | Text tasks |
| `mimo-v2.6-pro` | Alternative for tasks you want to compare with Flash | Text tasks |
| `glm-5.3-flash` | Lower token price than the GLM flagship; visual inputs | Text and image tasks |
| `glm-5.3` | Current GLM flagship | Text tasks |

A cheaper token rate does not guarantee a cheaper completed task. Keep the model
that produces correct results with acceptable time and cost for your work.
Premium speed variants are not included in this update.

## Existing MiMo users

Xiaomi's retirement notice, updated October 9, 2026, says requests to `mimo-v2.5-pro` and
`mimo-v2.5` are automatically routed to `mimo-v2.6-pro` and `mimo-v2.6-flash` from
**October 14, 2026** (the notice gives both 10:00 and 18:00 UTC+8), and that the old IDs become
invalid on **October 21, 2026 at 10:00 UTC+8 (02:00 UTC)**.
[Official retirement notice](https://mimo.mi.com/docs/en-US/updates/deprecate).

The unreleased source catalog no longer contains the two retiring routes, so the next build's
`/provider mimo` selects `mimo-v2.6-flash`. If your saved model choice names a removed route,
Nosis reports it; choose another with `nh model set <id>`. A session recorded on a removed route
cannot be resumed; start a new session instead. MiMo currently has no route advertised for image
input.

**v0.2.2 contains only the retiring MiMo routes,** so its MiMo requests fail after October 21,
2026.

**For v0.2.2 users:** no published Nosis build contains these new routes yet.
Keep your current installation and watch the [releases page](https://github.com/nosistech/nosis-harness/releases)
for a build that includes them. Rerunning v0.2.2 setup cannot add these routes.
If you need them before a release, build the published source using
[the contributor instructions](../CONTRIBUTING.md); there is no ready download
containing these changes yet. You can explicitly choose another supported provider meanwhile.

The commands below require a build containing the new catalog. After installing it, run this in your project:

```powershell
nh catalog migrate
```

Review the changes and backup location before confirming. Existing model choices
are not silently replaced. Custom catalogs require manual review. See
[catalog migration](CONFIGURATION.md#upgrade-your-project-catalog-v030-rc1-candidate).

To try the new Flash route explicitly:

```powershell
nh run "Explain the purpose of this project's README" --read-only --model mimo-v2.6-flash
```

The route uses your existing `mimo` credential entry. To make it your saved choice:

```powershell
nh model set mimo-v2.6-flash
```

Use `mimo-v2.6-pro` instead if that is the route you choose. New MiMo routes currently
advertise text only because the provider's new modality documentation is inconsistent.
This update does not claim verified image, video or audio integration for them.
[MiMo API contract](https://mimo.mi.com/docs/en-US/api/chat/openai-api).

## GLM and other providers

The new GLM routes use the ordinary API endpoint and `glm` credential entry.
Subscription coding-plan quota is a separate service. GLM 5.3 requires reasoning;
`--think none` becomes `low` for these routes: reasoning stays enabled and its
tokens can be billed. It is not a way to turn reasoning charges off.
Its default posture uses the provider's documented `max` effort; explicit
`--think high` requests less effort, and `--profile frugal` uses `low`. More effort
can consume more reasoning tokens. Inspect `nh profile --model glm-5.3` before a
long task; choosing a profile does not guarantee a final bill.
[GLM model contract](https://docs.z.ai/guides/llm/glm-5.3).

Existing DeepSeek Flash already calls the current canonical provider model. Existing
Kimi routes remain available. Their presence is not a promise of account eligibility,
rate-limit availability or permanent provider pricing. Catalog prices are estimates;
the provider's bill remains authoritative.

MiMo routes and Kimi K3 now use the documented `max_completion_tokens` request field
for their output ceiling. This includes reasoning tokens where the provider specifies
that behavior. It does not create a hard spending cap across calls or retries.

## Price and compatibility limits

The September 25 review checked all 16 catalog entries against current provider
documentation where available. It did not repeat live requests for every model.
Only the four additions above have the recorded September 22 text/tool smoke
checks; a catalog entry is not evidence of reliability on your task. Current
dynamic Kimi documentation did not independently reconfirm the older K2.7 Code
Highspeed rates, so check the provider's account pricing before choosing it.

DeepSeek's official peak schedule excludes Chinese public holidays. Nosis uses
weekday/time windows and can therefore overestimate on those holidays. A task
can also cross a price window after its initial quote.
[DeepSeek pricing](https://api-docs.deepseek.com/quick_start/pricing/).

Kimi K3 separately reports cache-write tokens. Nosis currently folds them into
the non-cache-read input bucket; its default five-minute write rate matches
ordinary input pricing. One-hour caching is not requested or priced by Nosis.
The receipt does not separately show cache writes or their lifetime.
[Kimi caching](https://platform.kimi.ai/docs/guide/context-caching).

Coding subscriptions may use different endpoints, credentials and entitlements.
For example, Kimi K2.8 Preview is listed in Kimi Code's service; it is not a drop-in
model ID for Nosis's existing Moonshot API route. Premium speed variants and new
modalities require explicit integration and testing before being advertised here.
[Kimi Code models](https://www.kimi.com/code/docs/en/kimi-code/models.html).