pi-models-access

The complete pi extension set for Alibaba's model lineup — Qwen 3.8 Max, Qwen 3.7 Max / Plus / Flash, Qwen 3.6 Plus / Flash, DeepSeek V4 Pro / Flash, Kimi K2.6 / K2.7, GLM-5.1 / 5.2, MiniMax M2.5, and the rest of the catalog — plus the official DeepSeek API provider (Completions / Anthropic / Responses API modes, dynamic model list, balance query, image understanding) and the Zhipu GLM provider (智谱 Coding Plan + API, dynamic model list, documented pricing). Native thinking-level support, both Anthropic- and OpenAI-shaped APIs (including the new OpenAI Responses API), International / China / US / Japan / Frankfurt endpoints, both Coding Plan subscriptions and pay-per-token Cloud keys.

Origin

This extension is a fork and extension of pi-alibaba-models by Francesco Frapporti at Fornace (the author field and the "From the same author" section below reflect the original work). We forked it and extended it with the official DeepSeek API provider (Completions / Anthropic / Responses API modes, dynamic model list, balance query, image understanding), the OpenAI Responses API shape for the Cloud provider, and renamed the package to pi-models-access.

Features

• Alibaba providers: subscription-based Model Studio Coding Plan and pay-per-token Alibaba Cloud (DashScope) — registered side by side, switch per chat from the model picker.
• DeepSeek official provider: deepseek-v4-flash / deepseek-v4-pro / deepseek-v4-flash-vision-exp against api.deepseek.com in three switchable API modes (/deepseek → API Format), with GET /models dynamic catalog, GET /user/balance (command + deepseek_balance tool), vision input on the exp model, and a footer balance line (DS ¥12.34). When no key is present the provider makes zero startup requests (no /models, no pricing-page sync).
• Zhipu GLM provider: overrides pi's built-in zai-coding-cn (智谱开放平台) and zai (Z.AI intl) providers, reusing their /login flow. Both Coding Plan (subscription quota) and API (pay-per-token) billing modes (/zhipu → Mode), GET /models dynamic catalog, documented CNY/USD pricing, and capability defaults (vision, context window, thinking) filled per model — GLM-5.3 included with its always-on thinking.
• Three API Shapes (Alibaba): Anthropic-compatible (/v1/messages) by default; OpenAI-compatible Chat Completions (/compatible-mode/v1) auto-selected for DeepSeek on the Anthropic format; the new OpenAI Responses API (/compatible-mode/v1/responses) selectable per-Cloud via /alibaba (DeepSeek-V4 included, Beijing/Singapore).
• Dual Provider Support (Alibaba): Both the subscription-based Model Studio Coding Plan and the pay-per-token Alibaba Cloud (DashScope) — registered side by side, switch per chat from the model picker.
• Five Regions: International (dashscope-intl.aliyuncs.com), China (dashscope.aliyuncs.com), US-Virginia (dashscope-us.aliyuncs.com), Japan-Tokyo and Germany-Frankfurt (workspace domains), plus Alibaba's recommended business-space (workspace) domains for Beijing/Singapore — switch with /alibaba, no re-login needed.
• Native Reasoning: First-class thinking-level support for every reasoning-capable model, including reasoning_effort levels (Qwen 3.8 Max: low/medium/xhigh; DeepSeek V4, GLM 5.x: high/max) on the OpenAI path.
• Vision Capable: Image input automatically enabled for VL models, Qwen 3.8 Max, and Qwen 3.x Plus variants.
• Live Catalog: Pulls the real /v1/models from DashScope on every login + the canonical Qwen-Code plan template. New models appear as Alibaba ships them — no extension update needed. The Cloud catalog now uses Alibaba's native GET /api/v1/models (real context windows, max output tokens, Reasoning/VU capability tags, input modality and pricing — no more id-based guessing where available), falling back to the compatible-mode endpoint on domains that don't expose it.
• Footer status lines: DeepSeek shows your account balance in the footer (DS ¥12.34, refreshed after each completed conversation round and on a 15-minute timer — toggle with /deepseek → Balance in status bar), only while the active model is a DeepSeek model (switching to another provider hides it, same as the Alibaba quota line); Alibaba Cloud shows the active model's rate-limit quota when the limits endpoint is reachable (Beijing workspace domain).

How to Use (Quickstart)

1. Install the extension (see below).
2. Restart pi to load the extension.
3. Type /login in your pi chat input.
4. Select your provider based on your account type:
   • Choose Plans > Alibaba Model Studio Coding Plan if you have a subscription (your token likely starts with sk-sp- or sk-tok-).
   • Choose Use an API key > Alibaba Cloud (API Key) if you use the pay-as-you-go DashScope service (your token likely starts with sk-).
5. Paste your token when prompted.
6. Open the model picker, select a model (e.g., Qwen 3.8 Max, Qwen 3.7 Plus, Qwen 3.6 Flash, or DeepSeek V4 Pro), and start chatting!

Install

# recommended
pi install pi-models-access

# explicit npm form (fallback if the bare name doesn't resolve)
pi install npm:pi-models-access

# or from GitHub
pi install git:github.com/blueye-y/pi-models-access

# or from a local checkout (development)
git clone https://github.com/blueye-y/pi-models-access
cd pi-models-access && pi install .

After install, restart pi. The extension registers two providers and a slash command on every boot.

Uninstall

pi remove only removes the package entry from settings.json["packages"] — it does not clean extension-private state (auth entries, config, model cache, enabled-model lists). For a clean uninstall:

1. /alibaba  →  "Reset all"      (wipes config, both auth entries, plan-models cache, alibaba-* enabledModels)
2. pi remove pi-models-access

If you've already run pi remove and want to clean leftovers manually:

rm -f ~/.pi/agent/alibaba-config.json ~/.pi/agent/alibaba-plan-models.cache.json
# then edit ~/.pi/agent/auth.json and remove the "alibaba-plan" / "alibaba-cloud" entries
# then edit ~/.pi/agent/settings.json and drop any "alibaba-*/..." or "dashscope/..." entries from enabledModels

Two providers

Provider id	Section in /login	Auth shape	Use it for
alibaba-plan	Plans	OAuth (paste token)	Model Studio Coding Plan subscription
alibaba-cloud	Use an API key	API key (paste via /login, or $DASHSCOPE_API_KEY)	Pay-per-token DashScope API

The Plan provider registers an oauth-shaped login (paste token) so it appears under Plans in /login; the Cloud provider is API-key-registered under “Use an API key”, and also works from the DASHSCOPE_API_KEY env var. Credentials live in ~/.pi/agent/auth.json under their respective keys; the Plan provider stores the chosen endpoints in the refresh field as JSON, and the Cloud provider stores its domain in ~/.pi/agent/alibaba-config.json.

Cloud without /login: the Cloud provider also reads the DASHSCOPE_API_KEY environment variable. If it's set, the extension fetches your live model catalog from it on startup — no /login needed. With no credential at all (no /login, no env var) the Cloud provider still shows up in /login → Use an API key via a single placeholder model, so you can sign in; your real catalog replaces it the moment a key is present.

Endpoints

Plan (default Singapore / Global):

• Anthropic-compat: https://token-plan.ap-southeast-1.maas.aliyuncs.com/apps/anthropic (pi appends /v1/messages)
• OpenAI-compat: https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1

Cloud (default International):

• Anthropic-compat: https://dashscope-intl.aliyuncs.com/apps/anthropic
• OpenAI-compat: https://dashscope-intl.aliyuncs.com/compatible-mode/v1 (Chat Completions and Responses)

Cloud — other regions (set via /alibaba → Cloud — Change Domain):

• China: https://dashscope.aliyuncs.com
• US (Virginia): https://dashscope-us.aliyuncs.com
• Beijing / Singapore / Tokyo / Frankfurt workspace domains: https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com, https://{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com, https://{WorkspaceId}.ap-northeast-1.maas.aliyuncs.com, https://{WorkspaceId}.eu-central-1.maas.aliyuncs.com — Alibaba recommends migrating to these for Beijing/Singapore; the old shared domains still work.

OpenAI Responses API (/compatible-mode/v1/responses): Alibaba's newest OpenAI-compatible surface, with built-in tools (web search, code interpreter, web extractor) and reasoning.effort thinking levels (none/minimal/low/medium/high). Select it per-Cloud via /alibaba → Cloud — Change API Format → OpenAI Responses. DeepSeek-V4 is supported on it too — but only in the Beijing and Singapore regions. When the format is Anthropic, DeepSeek models still fall back to Chat Completions (the Anthropic-compat path hangs for them).

Key prefix reference

Prefix	Provider	Where to obtain
sk-sp-	alibaba-plan	Model Studio Coding Plan console — Singapore / Global
sk-tok-	alibaba-plan	Model Studio Coding Plan console — alternate token format
sk-(other)	alibaba-cloud	DashScope API Keys console (per-token billing)

Consoles:

• International / Singapore Coding Plan: https://modelstudio-intl.console.alibabacloud.com/
• China Coding Plan: https://bailian.console.aliyun.com/
• DashScope (per-token): https://dashscope.console.aliyun.com/ or https://dashscope-intl.console.aliyun.com/

The login flow validates the prefix and offers to redirect you to the correct provider if you paste the wrong type.

Region table

Region	Plan host	Cloud host
International	token-plan.ap-southeast-1.maas.aliyuncs.com	dashscope-intl.aliyuncs.com
China	(region-specific host, paste via "Custom")	dashscope.aliyuncs.com
US (Virginia)	—	dashscope-us.aliyuncs.com
Hong Kong	—	cn-hongkong.dashscope.aliyuncs.com
Beijing (workspace)	{WorkspaceId}.cn-beijing.maas.aliyuncs.com	{WorkspaceId}.cn-beijing.maas.aliyuncs.com
Singapore (workspace)	{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com	{WorkspaceId}.ap-southeast-1.maas.aliyuncs.com
Japan (Tokyo)	—	{WorkspaceId}.ap-northeast-1.maas.aliyuncs.com
Germany (Frankfurt)	—	{WorkspaceId}.eu-central-1.maas.aliyuncs.com
US (workspace)	—	{WorkspaceId}.us-east-1.maas.aliyuncs.com
Custom	paste both base URLs at login	paste domain at login

The Beijing / Singapore workspace domains are the ones Alibaba recommends (and they're required for Japan/Frankfurt). /alibaba → Cloud — Change Domain asks for your business-space ID (业务空间详情 in the console) and builds the host for you.

Rate limits

/alibaba → Rate limits (Cloud) queries GET /api/v1/models/limits with your API key and prints each model's quota: request rate (req/s or req/min), token usage limit per period, and async queue/concurrency where applicable. It's a read-only best-effort view — the endpoint is currently only documented on the Beijing workspace domain, so on other domains it tells you to switch (/alibaba → Cloud — Change Domain) and retry. The view is filtered to the models in your current Cloud catalog so it stays focused. When the active model is an alibaba-cloud model, its rate limit is also shown in the footer (Ali qwen3.7-max 60 req/min, refreshed on a 15-minute timer; toggle with /alibaba → Status in footer).

Authorized-models filter

GET /api/v1/models/permissions returns the models your business space is authorized to call (authorization_scope=AUTHORIZED, action=INFERENCE). When the endpoint is reachable (Beijing workspace domain), the extension intersects it with the live catalog and hides models you don't have inference permission for — no more picking a model that 404s only at request time. It's on by default; disable any time via /alibaba → Cloud — Authorized-only Filter (or set cloudAuthorizedOnly: false in ~/.pi/agent/alibaba-config.json). If the fetch fails (non-Beijing domains) or the intersection would be empty, the full catalog is shown instead.

Studio plan models — dynamic source

The plan model list is fetched from the canonical Qwen Code template:

https://github.com/QwenLM/qwen-code/blob/main/packages/cli/src/constants/codingPlan.ts

Cached at ~/.pi/agent/alibaba-plan-models.cache.json for 4 hours. The live API is always the source of truth; on a failed fetch the extension falls back to the last-known-good on-disk cache (and, if there's no cache either, registers an empty list rather than crashing). Force a refresh from /alibaba → Refresh model lists.

The Cloud provider mirrors the live catalog from Alibaba's native GET /api/v1/models (paginated, capabilities=TG for text generation, then filtered for chat-capable ids) — Qwen 3.8 Max, Qwen 3.7 Max/Plus/Flash, Qwen 3.6 Plus/Flash, DeepSeek V4 Pro/Flash, Kimi K2.6/K2.7, GLM-5.1/5.2, MiniMax M2.5 etc. all surface automatically as Alibaba ships them, with real context windows, max-token limits and per-1M-token pricing from the API. Domains without /api/v1/models fall back to the compatible-mode endpoint (id-based inference). (DeepSeek v3.x line was retired by Alibaba in July 2026 — migrate to DeepSeek V4 or Qwen 3.x.)

Limitations & Known Issues

• DeepSeek Compatibility: The Anthropic-compatible path on the Alibaba Plan host often hangs or times out for DeepSeek models. To resolve this seamlessly, this extension automatically forces any model ID containing deepseek to use the OpenAI-completions endpoint instead.
• Model Availability (404s): The picker shows the catalog your account is actually authorized to call when the permissions endpoint is reachable (GET /api/v1/models/permissions, available on the Beijing workspace domain) — models you lack inference permission for are hidden instead of 404ing at request time. Toggle with /alibaba → Cloud — Authorized-only Filter. When the endpoint is unreachable (other domains), the full advertised catalog is shown and a model you can't access still errors only when you actually send a message.
• API Wrapper Quirks: Alibaba's native Anthropic compatibility layer can occasionally be strict or quirky with complex parallel tool calls. If you experience systemic parsing errors on DashScope, you can use the /alibaba command to switch your Cloud API format to "OpenAI Chat Completions" or "OpenAI Responses".
• Responses API: the OpenAI Responses surface supports DeepSeek-V4 only on Beijing/Singapore — on other regions those models fall back to Chat Completions, and not every model supports every built-in tool. If a model errors out, switch back to Chat Completions or Anthropic for that session.
• Dynamic Caching: Model lists are cached for 4 hours. If a new model drops and you don't see it, run /alibaba -> Refresh model lists.
• Inferred Context Windows: The /v1/models API returns only ids and names, so context windows are inferred from the model id. If a brand-new model shows the wrong size, fix it yourself with /alibaba → Context Window — Override (per model, or * for all) — no extension update needed.

/alibaba command reference

Choice	What it does
Status	Print Plan/Cloud login state, active endpoints, model count, cache age
Refresh model lists	Force-refetch Plan + Cloud catalogs and reload the extension
Re-login Plan	Wipe alibaba-plan from auth.json and reload (then run /login)
Re-login Cloud	Wipe alibaba-cloud from auth.json and reload (then run /login)
Plan — Change Endpoints	Override OpenAI / Anthropic base URLs
Cloud — Change Domain	International / China / US / Japan / Frankfurt / workspace domains / Custom
Cloud — Change API Format	Anthropic Messages / OpenAI Chat Completions / OpenAI Responses
Rate limits (Cloud)	Show per-model rate/usage quotas for the current API key (/api/v1/models/limits)
Status in footer — toggle	Show/hide the active Cloud model's rate-limit quota in the footer
Cloud — Authorized-only Filter	Toggle hiding catalog models the account isn't authorized to call (/api/v1/models/permissions)
Context Window — Override	Set the context-window shown on a model's card (per model, or * for all)
Reset all	Wipe all Alibaba state (config, both auth entries, plan-models cache)

Troubleshooting

• Model picker shows "No matching models" → run /login, pick the right Alibaba entry, paste your key. Models register only after a successful login (Cloud fetches its real model list at boot from the live key).
• sk-sp- accidentally pasted into the Cloud slot → run /alibaba → Re-login Cloud, then /login → Alibaba Model Studio Coding Plan and paste it there. (The login validators will also catch this and offer to redirect you.)
• DeepSeek hangs / times out → make sure you're on the latest version of this extension; it forces DeepSeek to OpenAI-compat. If you customised plan endpoints, verify the OpenAI URL ends in /compatible-mode/v1.
• Plan picker shows models that 404 at request time → your subscription tier may not include every advertised model. The picker shows whatever upstream advertises; the API tells you "model_not_found" only when you actually call it.
• /alibaba command doesn't appear → pi list should show pi-models-access (or whatever source you installed from) under "User packages". If absent, run pi install pi-models-access again and restart pi.

Files

Path	Purpose
~/.pi/agent/auth.json	Provider credentials (0600)
~/.pi/agent/alibaba-config.json	Alibaba endpoint / domain / format config
~/.pi/agent/alibaba-plan-models.cache.json	4 h plan-models cache
~/.pi/agent/alibaba-cloud-models.cache.json	4 h cloud-models cache
~/.pi/agent/deepseek-config.json	DeepSeek API format / base URL config
~/.pi/agent/deepseek-models.cache.json	DeepSeek GET /models cache
~/.pi/agent/deepseek-prices.cache.json	Last-synced official pricing (both peak/off-peak rate sets)

DeepSeek official API provider (/deepseek)

Since v1.1.0 the package also ships a DeepSeek official API extension (extensions/deepseek.ts). It follows the official documentation at api-docs.deepseek.com/zh-cn and implements the full capability surface:

• Three API modes, switchable per session via /deepseek → API Format:
  • openai-completions — POST https://api.deepseek.com/chat/completions (default)
  • anthropic-messages — POST https://api.deepseek.com/anthropic/v1/messages
  • openai-responses — POST https://api.deepseek.com/responses
• Thinking mode per the official docs — {"thinking":{"type":"enabled/disabled"}}
  • reasoning_effort (OpenAI), output_config.effort (Anthropic), reasoning.effort (Responses). The documented effort mapping is baked in: low→low, medium→high, high→high, xhigh→high, max→max; thinking defaults on at high like the raw API.
• Dynamic model list — GET /models at startup and on /deepseek → Refresh model list, with an offline cache fallback (~/.pi/agent/deepseek-models.cache.json).
• Balance query — GET /user/balance via /deepseek → Balance or the deepseek_balance tool (is_available + per-currency total/granted/topped-up balances, CNY/USD), plus a footer balance line (DS ¥12.34) refreshed after each completed conversation round and on a 15-minute timer (/deepseek → Balance in status bar — toggle).
• Price sync — model costs are parsed from the official pricing page (api-docs.deepseek.com/quick_start/pricing) once per extension load, caching both peak and off-peak USD rates (~/.pi/agent/deepseek-prices.cache.json). The active rate set follows the UTC clock (peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, exactly 2× off-peak), re-registering the provider at each boundary so the footer $ cost stays accurate. Offline, the last-synced cache (then a built-in table) is used.
• Image understanding — deepseek-v4-flash-vision-exp (JPEG/PNG/GIF/WebP) is included with input: ["text", "image"]; toggle it via /deepseek → Vision model.
• Model catalog — deepseek-v4-flash, deepseek-v4-pro, deepseek-v4-flash-vision-exp (1M context / 384K max output), plus any ids returned by /models — new models appear without an extension update.

Quickstart

1. pi install pi-models-access (or run from a checkout) and restart pi.
2. /login → Use an API key → DeepSeek and paste your key from platform.deepseek.com/api_keys — or just export DEEPSEEK_API_KEY (the provider is registered with apiKey: "$DEEPSEEK_API_KEY").
3. Open the model picker and pick DeepSeek V4 Pro / DeepSeek V4 Flash / DeepSeek V4 Flash Vision (exp).
4. /deepseek → API Format to switch between the Completions, Anthropic and Responses API modes (reloads automatically).

/deepseek command reference

Menu item	Action
Status	Auth state, API format, base URLs, model count, cache age, price sync state, current peak/off-peak period
Refresh model list	Force GET /models and re-register the provider
API Format	Switch openai-completions / anthropic-messages / openai-responses
Balance	GET /user/balance (is_available + CNY/USD totals), then refreshes the footer line
Balance in status bar — toggle	Show/hide the account balance in the footer
Re-login	Wipe the deepseek auth entry and re-run /login
Base URL — override	Custom OpenAI/Anthropic base URLs (proxies/gateways)
Vision model — toggle	Show/hide deepseek-v4-flash-vision-exp
Context Window — override	Per-model (or *) context-window override
Reset all	Wipe config, models cache and auth entry

Notes

• The Anthropic mode uses a custom streaming implementation that sends exactly what the docs specify for that endpoint (thinking.type + output_config.effort), so thinking levels work on all three formats.
• deepseek_balance is a custom tool the LLM can call on its own.
• Context caching (KV cache) is automatic server-side per the docs — cache-hit tokens are surfaced in usage as cacheRead.

Zhipu GLM provider (/zhipu)

The package also ships a Zhipu (智谱) GLM extension (extensions/zhipu.ts) that overrides pi's built-in zai-coding-cn (智谱开放平台, open.bigmodel.cn) and zai (Z.AI international, api.z.ai) providers, so the built-in /login flow keeps working unchanged — the auth.json entries and $ZAI_CODING_CN_API_KEY / $ZAI_API_KEY env vars are inherited from the built-ins (same trick the DeepSeek extension uses to reuse pi's default DeepSeek login).

• Two billing modes, switchable via /zhipu → Mode:
  • Coding Plan (subscription quota, default): https://open.bigmodel.cn/api/coding/paas/v4
  • API (pay-per-token): https://open.bigmodel.cn/api/paas/v4
  • intl equivalents on api.z.ai. Both use the OpenAI Chat Completions shape (thinking format zai, zaiToolStream, max_tokens) — the same shape as the built-ins, so no custom streaming is needed.
• Dynamic model list — GET /models at startup and on /zhipu → Refresh model lists, with an offline cache fallback (~/.pi/agent/zhipu-models.cache.json) and the built-in catalog as last resort.
• Live model specs — context window, max output, vision and thinking are parsed from the official docs pages (docs.bigmodel.cn serves every model page as markdown; discovery goes through the model-overview index), cached for 7 days (~/.pi/agent/zhipu-specs.cache.json) and refreshed via /zhipu → Refresh model lists. Synced specs override the built-in catalog, so spec changes land without an extension update. Only runs when logged in.
• Documented pricing — CNY per 1M tokens (USD for intl): GLM-5.3/5.2 ¥8 in / ¥28 out / ¥2 cache hit, GLM-5 ¥4 / ¥18 / ¥1, GLM-5.3-Flash ¥0.8 / ¥2.8 / ¥0.23. Unknown ids fall back to default rates. (智谱's pricing page is a JS-rendered console page with no public static source, so prices come from the documented table rather than a live page parse.)
• Capability defaults filled per model id — context window, max output, vision and thinking: GLM-5.3 is 1M ctx with thinking always on (its off thinking level is hidden because the API rejects thinking.type: disabled); GLM-5.3-Flash / 5V / 4.6V / 4.1V models are flagged vision-capable; flash/free models can be hidden via showFlash in zhipu-config.json.

Quickstart

1. pi install pi-models-access (or run from a checkout) and restart pi.
2. /login → Use an API key → Z.AI Coding CN (China) or Z.AI (intl) and paste your Coding Plan token or API key from the 智谱 console — or export ZAI_CODING_CN_API_KEY / ZAI_API_KEY.
3. Open the model picker and pick e.g. GLM-5.3.
4. /zhipu → Mode to switch between Coding Plan and API billing.

/zhipu command reference

Menu item	Action
Status	Mode, flash toggle, spec-sync state, per-region auth, base URL, model count, cache age
Refresh model lists	Force GET /models for both regions and re-register the providers
Mode — Coding Plan / API	Switch the billing endpoint
Re-login	Wipe the zai/zai-coding-cn auth entries and re-run /login
Base URL — override	Custom OpenAI-compatible base URL per region (proxies/gateways)
Context Window — override	Per-model (or *) context-window override
Reset all	Wipe config, models cache and auth entries

Note — there is no account-balance/quota command for Zhipu: the platform has no documented API-key balance endpoint (the console balance page is session-authenticated).

From the same author

By Francesco Frapporti at Fornace.

• pi-bench — LLM benchmark toolkit for pi. Probes every available model to find the fastest and cheapest.
• pi-recap — Always-visible session recap panel for pi. Uses pi-bench data to pick the fastest summarization model.
• pi-banana — Generate and edit images inside pi using Google Nano Banana. All package banners in this ecosystem were created with pi-banana.
• pi-notte-theme — Notte: a true-dark pi theme where darkness has color and text glows like terminal phosphor.