BlockRunAI · VickyXAI · Sep 29, 2026 · Sep 29, 2026
diff --git a/README.md b/README.md
@@ -8,7 +8,7 @@
 > request settle itself over x402 — on **Solana or Base**. Every client takes
 > either credential in the same first argument.
 >
-> 🆓 **Includes 8 fully-free NVIDIA-hosted models** — DeepSeek V4 Flash (1M context), Nemotron Nano Omni (vision), Qwen3 Next + Coder, Llama 4 Maverick, Mistral Small 4, plus `gpt-oss-120b/20b` (hidden from `/v1/models` but direct calls still work). Zero USDC, no rate-limit gimmicks. Use `routing_profile="free"` or call any `nvidia/*` model directly.
+> 🆓 **Includes <!-- br:models.free -->6<!-- /br:models.free --> free models** — Nemotron 3.5 Lightning (1M context), Nemotron 3 Ultra 550B, Nemotron 3 Nano Omni, Llama 3.2 11B Vision, Cohere North Mini Code and Poolside Laguna XS 2.1. Zero USDC, no rate-limit gimmicks. Use `routing_profile="free"` or call any of them directly.
 
 [![PyPI](https://img.shields.io/pypi/v/blockrun-llm.svg)](https://pypi.org/project/blockrun-llm/)
 [![License: MIT](https://img.shields.io/badge/License-MIT-green.svg)](LICENSE)
@@ -62,7 +62,7 @@ reports which rail you ended up on.
 ### Try It Free (No Balance Required)
 
 Want to kick the tires before topping up or funding a wallet? Route to
-BlockRun's free NVIDIA tier — it settles $0 on both rails, so an unfunded wallet
+BlockRun's free tier — it settles $0 on both rails, so an unfunded wallet
 or a $0 credit account is enough:
 
 ```python
@@ -71,31 +71,30 @@ from blockrun_llm import LLMClient
 client = LLMClient()  # a credential is still needed; a balance is not
 
 # Option 1: call a free model directly
-response = client.chat("nvidia/step-3.7-flash", "Explain x402 in 1 sentence")
+response = client.chat("nvidia/nemotron-3.5-lightning", "Explain x402 in 1 sentence")
 
 # Option 2: let the smart router pick the best free model per request
 result = client.smart_chat("What is 2+2?", routing_profile="free")
-print(result.model)     # e.g. 'nvidia/step-3.7-flash' (cheapest capable for SIMPLE tier)
+print(result.model)     # e.g. 'nvidia/nemotron-3-nano-omni-30b-a3b-reasoning' (free SIMPLE tier)
 print(result.response)  # '4'
 ```
 
-**Available free models** (input + output both $0, all NVIDIA-hosted):
+**Available free models** (input + output both $0; the live list is `GET /v1/models` filtered on $0 pricing):
 
 | Model ID | Context | Best For |
 |----------|---------|----------|
-| `nvidia/step-3.7-flash` | 131K | Fast general-purpose chat + reasoning |
-| `nvidia/mistral-nemotron` | 131K | Fast free Mistral (Mistral × NVIDIA) |
-| `nvidia/nemotron-3-nano-omni-30b-a3b-reasoning` | 256K | Only vision-capable free model — text + images + video (≤2 min) + audio (≤1 hr) |
-| `nvidia/nemotron-nano-9b-v2` | 131K | Compact fast chat |
-| `nvidia/nemotron-nano-12b-v2-vl` | 131K | Compact vision |
-| `nvidia/gpt-oss-120b` | 128K | OpenAI open-weight 120B — the free workhorse. Hidden from `/v1/models` (so SmartChat won't auto-pick it) but direct calls still work |
-| `nvidia/gpt-oss-20b` | 128K | OpenAI open-weight 20B — 155 tok/s. Hidden from `/v1/models` but direct calls still work |
+| `nvidia/nemotron-3.5-lightning` | 1M | The free default — thinking-mode reasoning, longest free context |
+| `nvidia/nemotron-3-ultra-550b` | 1M | Largest free model (550B / 55B active MoE) — strongest, but slower |
+| `nvidia/nemotron-3-nano-omni-30b-a3b-reasoning` | 256K | Fast general chat + reasoning (30B-A3B) |
+| `nvidia/llama-3.2-11b-vision` | 128K | The free Meta Llama |
+| `cohere/north-mini-code` | 256K | Coding, sub-second responses |
+| `poolside/laguna-xs-2.1` | 131K | Coding, ~161 tok/s |
 
-> Need V4-Pro-class reasoning? Use the paid `deepseek/deepseek-v4-pro` ($0.435/$0.87 — the 75% launch promo became the permanent list price after 2026-05-31) — `nvidia/deepseek-v4-pro` is hidden because NVIDIA's NIM deployment is hung; backend MODEL_REDIRECTS forwards calls to V4 Flash.
+> Two of these (`nemotron-3-nano-omni` and `llama-3.2-11b-vision`) are catalogued as vision-capable, but image input did not hold up on real probes — both return HTTP 200 with a wrong answer. Send images to a paid vision model.
 
-> **Privacy note for `gpt-oss-120b/20b`**: NVIDIA's free build.nvidia.com tier reserves the right to use prompts/outputs for service improvement. The models are hidden from `/v1/models` so SmartChat won't auto-route to them, but direct calls still work — use them only when prompts contain no sensitive data.
+> Need V4-Pro-class reasoning? Use the paid `deepseek/deepseek-v4-pro` ($0.435/$0.87 — the 75% launch promo became the permanent list price after 2026-05-31) — `nvidia/deepseek-v4-pro` is hidden because NVIDIA's NIM deployment is hung; backend MODEL_REDIRECTS forwards calls to V4 Flash.
 
-> **Retired**: NVIDIA has EOL'd (HTTP 410) most of its early free lineup — the free DeepSeek family (last: `nvidia/deepseek-v4-flash`, 2026-08-12), `llama-4-maverick`, the qwen3 SKUs, free Mistral small/large, and more. The gateway auto-redirects pinned callers to a healthy free model, so old model IDs still return 200.
+> **Retired**: NVIDIA has EOL'd most of its early free lineup — the free DeepSeek family (last: `nvidia/deepseek-v4-flash`, 2026-08-12), `step-3.7-flash`, `mistral-nemotron`, `nemotron-nano-9b-v2` / `-12b-v2-vl` (2026-08-30), `gpt-oss-120b/20b` (2026-09-03), `nemotron-3-nano-30b` (2026-09-08), `llama-4-maverick`, the qwen3 SKUs, free Mistral small/large, and more. The gateway auto-redirects pinned callers to a live model, so old model IDs still return 200 — from a different model.
 
 ## Solana Support
 
@@ -228,7 +227,7 @@ print(decision.reasoning)   # human-readable explanation of the pick
 
 | Profile | Description | Best For |
 |---------|-------------|----------|
-| `free` | NVIDIA free tier — smart-routes across the <!-- br:models.free -->6<!-- /br:models.free --> $0 models (Step 3.7 Flash, Mistral Nemotron, Nemotron Nano Omni / 9B / 12B VL) | Zero-cost testing, dev, prod |
+| `free` | Free tier — smart-routes across the $0 models (Nemotron 3 Nano Omni, Nemotron 3.5 Lightning, Llama 3.2 11B Vision, Cohere North Mini Code, Poolside Laguna XS 2.1). Nemotron 3 Ultra 550B is free too, by direct call. | Zero-cost testing, dev, prod |
 | `eco` | Cheapest capable model per tier | Cost-sensitive production |
 | `auto` | Best balance of cost/quality (default) | General use |
 | `premium` | Top-tier models (Anthropic, OpenAI, Moonshot) | Quality-critical tasks |
@@ -366,8 +365,8 @@ print(link["url"])  # open https://pay.coinbase.com/... to buy USDC on Base
 # (b) Transfer existing Base USDC to your wallet address
 print(client.get_wallet_address())  # send USDC on Base to this 0x… address
 
-# (c) Skip funding entirely — the free NVIDIA models cost $0
-client.chat("nvidia/step-3.7-flash", "Hello!")  # routing_profile="free" also works
+# (c) Skip funding entirely — the free models cost $0
+client.chat("nvidia/nemotron-3.5-lightning", "Hello!")  # routing_profile="free" also works
 ```
 
 `$5` of USDC covers thousands of paid requests. Check your balance any time:
@@ -516,24 +515,22 @@ glm-5 and glm-5-turbo on 2026-06-06) — the whole family now bills per-token.
 
 ### NVIDIA (Free & Hosted)
 
-Free tier refreshed 2026-08-12. NVIDIA has retired (HTTP 410 end-of-life)
-the entire free DeepSeek family — `nvidia/deepseek-v4-flash` was the last to
-go — along with `llama-4-maverick`, `qwen3-coder-480b`, the free Mistral
-small/large SKUs, and others. Retired models stay callable by ID: the gateway
-auto-redirects them to a healthy free model, so pinned callers still get a
-200. `nvidia/gpt-oss-120b` and `nvidia/gpt-oss-20b` remain callable by direct
-ID but are hidden from `/v1/models` over the NVIDIA free tier's
-prompt-retention terms (so SmartChat won't auto-pick them). The live list is
-`GET /v1/models` filtered on the free flag.
+Free tier checked against `/v1/models` on 2026-09-29. NVIDIA has retired
+(HTTP 410 end-of-life, or deprovisioned) most of its earlier free lineup — the
+free DeepSeek family, `step-3.7-flash`, the `nemotron-nano` v2 SKUs,
+`gpt-oss-120b/20b` (2026-09-03) and `nemotron-3-nano-30b` (2026-09-08) among
+them. Retired models stay callable by ID: the gateway auto-redirects them to a
+live model, so pinned callers still get a 200. The free tier also includes two
+non-NVIDIA models, `cohere/north-mini-code` and `poolside/laguna-xs-2.1` (see
+[Try It Free](#try-it-free-no-balance-required)). The live list is
+`GET /v1/models` filtered on $0 pricing.
 
 | Model | Input Price | Output Price | Context | Notes |
 |-------|-------------|--------------|---------|-------|
-| `nvidia/step-3.7-flash` | **FREE** | **FREE** | 131K | Fast general-purpose chat + reasoning |
-| `nvidia/mistral-nemotron` | **FREE** | **FREE** | 131K | Fast free Mistral |
-| `nvidia/nemotron-3-nano-omni-30b-a3b-reasoning` | **FREE** | **FREE** | 256K | First vision-capable free model — RGB images, mp4 video |
-| `nvidia/nemotron-nano-9b-v2` | **FREE** | **FREE** | 131K | Compact fast chat |
-| `nvidia/gpt-oss-120b` | **FREE** | **FREE** | 128K | OpenAI open-weight 120B — 123 tok/s. Hidden from `/v1/models`; direct calls work |
-| `nvidia/gpt-oss-20b` | **FREE** | **FREE** | 128K | OpenAI open-weight 20B — 155 tok/s. Hidden from `/v1/models`; direct calls work |
+| `nvidia/nemotron-3.5-lightning` | **FREE** | **FREE** | 1M | Free default — thinking-mode reasoning |
+| `nvidia/nemotron-3-ultra-550b` | **FREE** | **FREE** | 1M | Largest free model (550B / 55B active MoE) |
+| `nvidia/nemotron-3-nano-omni-30b-a3b-reasoning` | **FREE** | **FREE** | 256K | Fast general chat + reasoning |
+| `nvidia/llama-3.2-11b-vision` | **FREE** | **FREE** | 128K | Meta Llama 3.2 11B |
 | `moonshot/kimi-k2.5` | $0.60/M | $3.00/M | 262K | Kimi K2.5 direct from Moonshot (replaces `nvidia/kimi-k2.5`) |
 | `moonshot/kimi-k2.6` | $0.95/M | $4.00/M | 256K | Moonshot flagship (vision + reasoning_content) |
 
@@ -1856,7 +1853,7 @@ When you make an API call, the SDK automatically handles x402 payment. It signs
 Router Core is BlockRun's built-in routing engine — shared with the TypeScript SDK and the gateway, so the same request routes the same way everywhere. It scores your request across <!-- br:clawrouter.dimensions -->15<!-- /br:clawrouter.dimensions --> dimensions, drops every model that can't actually handle it (context, output length, tools, vision), then picks the cheapest capable one and keeps the rest as a fallback chain. Routing happens locally in under 1ms and makes no extra model call. It can save up to <!-- br:savings.autoVsBaselinePct -->84<!-- /br:savings.autoVsBaselinePct -->% on LLM costs compared to using premium models for every request.
 
 ### How much does it cost?
-Pay only for what you use. Prices start at **FREE** (11 NVIDIA-hosted models). Paid models start at $0.10/M tokens. There are no minimums, subscriptions, or monthly fees. $5 in USDC gets you thousands of requests.
+Pay only for what you use. Prices start at **FREE** (<!-- br:models.free -->6<!-- /br:models.free --> free models). Paid models start at $0.10/M tokens. There are no minimums, subscriptions, or monthly fees. $5 in USDC gets you thousands of requests.
 
 ### Can I use it with Solana?
 Yes. Install with `pip install blockrun-llm[solana]` and use `SolanaLLMClient` instead of `LLMClient`. Same API, different payment chain.

diff --git a/blockrun_llm/client.py b/blockrun_llm/client.py
@@ -1005,7 +1005,7 @@ def chat_completion_stream(
         request returns 402, the SDK signs an EIP-712 payment locally, then
         re-issues the request with ``stream=true`` and the
         ``PAYMENT-SIGNATURE`` header. Free models (e.g.
-        ``nvidia/deepseek-v4-flash``) skip the 402 and stream directly.
+        ``nvidia/nemotron-3.5-lightning``) skip the 402 and stream directly.
 
         Fallback semantics
         ------------------
@@ -1019,9 +1019,9 @@ def chat_completion_stream(
         Example::
 
             for chunk in client.chat_completion_stream(
-                "nvidia/deepseek-v4-flash",
+                "nvidia/nemotron-3.5-lightning",
                 [{"role": "user", "content": "Hello"}],
-                fallback_models=["nvidia/llama-4-maverick"],
+                fallback_models=["nvidia/nemotron-3-nano-omni-30b-a3b-reasoning"],
             ):
                 delta = chunk.choices[0].delta
                 if delta.content:

diff --git a/blockrun_llm/router_adapter.py b/blockrun_llm/router_adapter.py
@@ -73,17 +73,27 @@ class ResolvedRoutingDecision(RoutingDecision, total=False):
 #: ``step-3.7-flash``, ``nemotron-nano-9b-v2``, ``nemotron-nano-12b-v2-vl`` and
 #: ``mistral-nemotron`` (retired upstream 2026-08-30), plus ``nemotron-3-ultra-550b``
 #: and ``nemotron-3-nano-omni-30b-a3b-reasoning`` — the latter two still list at
-#: $0 in ``/v1/models`` but both answer as ``nemotron-3-nano-30b``.
+#: $0 in ``/v1/models`` but both answered as ``nemotron-3-nano-30b``.
+#:
+#: 2026-09-29: ``nemotron-3-nano-30b`` itself is gone. NVIDIA deprovisioned it for
+#: blockrun's account on 2026-09-08, it left ``/v1/models``, and the gateway now
+#: redirects it to nano-omni. Every slot it held goes to nano-omni: that is the
+#: model those slots were actually being served by, it is the same family and size
+#: (30B-A3B), and with nano-30b gone the reason it was excluded above no longer
+#: applies (a model-echo probe that day got nano-omni's own NIM deployment back).
+#: The router adapter was already dropping nano-30b at runtime, because it is not
+#: in the catalog, so SIMPLE had been opening on its first fallback.
 #:
 #: The table is no longer NVIDIA-only: ``cohere/north-mini-code`` and
 #: ``poolside/laguna-xs-2.1`` serve at $0 and carry the free coding load.
 #: ``gpt-oss-120b/20b`` stay out — proxy-only ids with no catalog price, under
 #: the NVIDIA free tier's prompt-retention policy.
 FREE_TIERS: dict[str, TierConfig] = {
     "SIMPLE": {
-        # Fastest free model (~121 tok/s), and latency is the only axis that
-        # separates free rungs — they all cost $0.
-        "primary": "nvidia/nemotron-3-nano-30b",  # 131K ctx
+        # Was nemotron-3-nano-30b, the fastest free model, until its 2026-09-08
+        # delisting. Latency is the only axis that separates free rungs (they
+        # all cost $0); nano-omni answered in 1.4s on blockrun's direct probe.
+        "primary": "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",  # 256K ctx
         "fallback": [
             "nvidia/nemotron-3.5-lightning",
             "nvidia/llama-3.2-11b-vision",
@@ -93,7 +103,7 @@ class ResolvedRoutingDecision(RoutingDecision, total=False):
     "MEDIUM": {
         "primary": "nvidia/nemotron-3.5-lightning",  # 1M ctx — free tier flagship
         "fallback": [
-            "nvidia/nemotron-3-nano-30b",
+            "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
             "poolside/laguna-xs-2.1",
             "cohere/north-mini-code",
         ],
@@ -104,14 +114,14 @@ class ResolvedRoutingDecision(RoutingDecision, total=False):
         "primary": "nvidia/nemotron-3.5-lightning",
         "fallback": [
             "cohere/north-mini-code",  # 256K ctx
-            "nvidia/nemotron-3-nano-30b",
+            "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
             "nvidia/llama-3.2-11b-vision",  # free vision
         ],
     },
     "REASONING": {
         "primary": "nvidia/nemotron-3.5-lightning",
         "fallback": [
-            "nvidia/nemotron-3-nano-30b",
+            "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
             "cohere/north-mini-code",
             "poolside/laguna-xs-2.1",
         ],

diff --git a/blockrun_llm/router_core/__init__.py b/blockrun_llm/router_core/__init__.py
@@ -2,7 +2,7 @@
 Router Core — deterministic, constraint-first model routing.
 
 Python port of `@blockrun/router-core <https://github.com/BlockRunAI/router-core>`_
-(upstream commit ``5ee7c23``), the same routing engine the TypeScript SDK and
+(upstream commit ``e38958b``), the same routing engine the TypeScript SDK and
 the BlockRun gateway use. The package is deliberately product-neutral: task
 classification, hard capability filtering, portfolio scoring, ordered
 fallbacks, and routing configuration. It contains no wallet, gateway client,

diff --git a/blockrun_llm/router_core/config.py b/blockrun_llm/router_core/config.py
@@ -2,7 +2,7 @@
 Default Routing Config
 
 Python port of ``@blockrun/router-core`` ``config.ts`` (upstream commit
-``5ee7c23``, 2026-08-30 — the same pin the TypeScript SDK
+``e38958b``, 2026-09-29 — the same pin the TypeScript SDK
 bundles). V3.5: every tier rung is a public catalog id.
 
 All routing parameters as a module constant. Hosts override by passing their
@@ -1156,7 +1156,12 @@
         "SIMPLE": {
             "primary": "nvidia/nemotron-3.5-lightning",  # FREE — NVIDIA free tier flagship, 1M ctx
             "fallback": [
-                "nvidia/nemotron-3-nano-30b",  # FREE — fastest free model (~121 tok/s)
+                # Was nvidia/nemotron-3-nano-30b until blockrun delisted it on 2026-09-08
+                # (NVIDIA deprovisioned it for the account — a structured per-account
+                # 404). nano-omni is blockrun's own redirect target for that id, so this
+                # follows the same rule as the head: the router and the gateway name the
+                # same model.
+                "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",  # FREE — nano-30b's successor, 256K ctx
                 # The free head keeps rotting with NVIDIA's hosting (deepseek-v4-flash
                 # 410 2026-08-12, seed-oss-36b 410 2026-08-03, gpt-oss-120b/20b 400
                 # 2026-08-21, and on 2026-08-30 FOUR of the five visible free models at

diff --git a/blockrun_llm/router_core/model_capabilities.py b/blockrun_llm/router_core/model_capabilities.py
@@ -188,18 +188,18 @@
             "supports_tools": False,
             "supports_vision": True,
         },
-        # supportsTools: not probed — fails closed
-        "nvidia/nemotron-3-nano-30b": {
-            "context_window": 131_072,
-            "max_output_tokens": 16_384,
-            "supports_tools": False,
-            "supports_vision": False,
-        },
+        # override: The catalog tags this model "vision", but a correctly sized probe does not
+        # hold up (ClawRouter, 2026-08-31): a 64x64 solid-red PNG was named correctly 1 of 4
+        # times on Base, and on Solana the image was silently dropped and a text model answered
+        # "white". An HTTP 200 with a confident wrong answer gives the caller nothing to branch
+        # on, so image turns must not be routed here. It is ecoTiers.SIMPLE.fallback[0] since
+        # nemotron-3-nano-30b was delisted (2026-09-08); remove once a probe of this size comes
+        # back right on both chains.
         "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning": {
             "context_window": 256_000,
             "max_output_tokens": 16_384,
             "supports_tools": False,
-            "supports_vision": True,
+            "supports_vision": False,
         },
         # supportsTools: not probed — fails closed
         "nvidia/nemotron-3-ultra-550b": {

diff --git a/tests/unit/router_core_decisions.snapshot.json b/tests/unit/router_core_decisions.snapshot.json
@@ -969,7 +969,7 @@
   "agenticScore": 0,
   "candidates": [
    "nvidia/nemotron-3.5-lightning",
-   "nvidia/nemotron-3-nano-30b",
+   "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
    "google/gemini-2.5-flash-lite",
    "zai/glm-5.3-flash",
    "openai/gpt-5.6-luna",