Upload 5 files

README.md

@@ -35,22 +35,26 @@ itself.
 
 To run it (locally or as a Space), set the `HF_TOKEN` secret/env var to a
 Hugging Face access token with Inference API permission. Optionally set
-`HF_AGENT_MODEL` to override the default model
-(`meta-llama/Llama-3.1-8B-Instruct`), and `HF_AGENT_PROVIDER` to override the
-default Inference Provider (`deepinfra`) — pinned explicitly because the
-router's `auto` selection can land on a provider that doesn't support tool
-calling for a given model (HTTP 422 `UNSUPPORTED_OPENAI_PARAMS`), and not every
-provider hosts every model (HTTP 4xx "Model ... is not supported by provider
-..."). Check `https://router.huggingface.co/v1/models/<repo_id>` to see which
-providers host a given model and whether they support tools before changing
-either variable.
+`HF_AGENT_MODEL` to override the default model (`Qwen/Qwen3-14B`), and
+`HF_AGENT_PROVIDER` to override the default Inference Provider (`nscale`) —
+pinned explicitly because the router's `auto` selection can land on a
+provider that doesn't support tool calling for a given model (HTTP 422
+`UNSUPPORTED_OPENAI_PARAMS`), and not every provider hosts every model (HTTP
+4xx "Model ... is not supported by provider ..."). Check
+`https://router.huggingface.co/v1/models/<repo_id>` to see which providers
+host a given model and whether they support tools before changing either
+variable.
 
-The default 8B model was chosen for cost, not accuracy: the HF free tier only
-includes $0.10/month of Inference Provider credits, and a 70B model burns
-through that almost instantly across a full GAIA run. If you have more budget
-(pre-paid credits or PRO), switch to a larger model such as
-`meta-llama/Llama-3.3-70B-Instruct` on the `together` provider for more
-reliable reasoning on harder questions.
+The default model is a budget pick, not a maximal-accuracy one: the HF free
+tier only includes $0.10/month of Inference Provider credits, and a 70B+
+model burns through that in a couple of questions. `Qwen/Qwen3-14B` on
+`nscale` costs only a few cents for a full ~20-question GAIA run while
+reasoning noticeably better than the 8B model it replaced. For an even
+smaller footprint (more runs per month out of the free allowance), set
+`HF_AGENT_MODEL=meta-llama/Llama-3.1-8B-Instruct` and
+`HF_AGENT_PROVIDER=deepinfra`. If you have real budget (pre-paid credits or
+PRO), go bigger with something like `meta-llama/Llama-3.3-70B-Instruct` on
+`together` for more reliable reasoning on harder questions.
 
 ### Using Claude instead
 

agent.py

@@ -34,18 +34,20 @@ DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
 LLM_PROVIDER = os.getenv("LLM_PROVIDER", "huggingface")
 
 # Any HF Inference API model that supports tool calling works here. Override via the
-# HF_AGENT_MODEL env var without touching code. Llama-3.1-8B-Instruct is ~15x cheaper per
-# token than the 70B default it replaced - picked to fit the free tier's tiny $0.10/month
-# Inference Provider credit allowance; swap to a larger model (e.g.
-# meta-llama/Llama-3.3-70B-Instruct on the "together" provider) once you have more budget
-# for better reasoning on harder GAIA questions.
-HF_MODEL_REPO_ID = os.getenv("HF_AGENT_MODEL", "meta-llama/Llama-3.1-8B-Instruct")
+# HF_AGENT_MODEL env var without touching code. Qwen3-14B gives noticeably better reasoning
+# than the 8B model it replaced (meta-llama/Llama-3.1-8B-Instruct) for a modest bite out of
+# the free tier's tiny $0.10/month Inference Provider credit allowance - on the "nscale"
+# provider below, a full ~20-question GAIA run costs only a few cents. Drop back to
+# Llama-3.1-8B-Instruct (paired with the "deepinfra" provider) if you need more runs per
+# month out of the free allowance, or go bigger (e.g. meta-llama/Llama-3.3-70B-Instruct on
+# "together") once you have paid credits.
+HF_MODEL_REPO_ID = os.getenv("HF_AGENT_MODEL", "Qwen/Qwen3-14B")
 
 # Left unset ("auto"), the HF router picks whichever provider is cheapest/fastest for the
 # model, and some of those don't support the `tools` param at all (HTTP 422
 # UNSUPPORTED_OPENAI_PARAMS). Pin one that reliably supports tool calling; override via
 # HF_AGENT_PROVIDER if it doesn't happen to host HF_AGENT_MODEL.
-HF_MODEL_PROVIDER = os.getenv("HF_AGENT_PROVIDER", "deepinfra")
+HF_MODEL_PROVIDER = os.getenv("HF_AGENT_PROVIDER", "nscale")
 
 # Cheapest current Claude model - a good fit for an agent that makes many tool-calling
 # round trips per question. Override via ANTHROPIC_MODEL (e.g. "claude-sonnet-5" for