[docs] Announce FastH3 on consumer hardware and FastH3 Trim

[aryan5v]

Summary

Announce FastH3 V2 on a single consumer machine (RTX 5090, RTX 4090, RTX PRO 6000, DGX Spark, Apple Silicon) and the experimental FastH3 Trim, with links to the models and the launch blog.

Code: #1919 (RTX) and #1920 (DGX Spark and Apple Silicon). Companion blog: hao-ai-lab/hao-ai-lab.github.io#108.

Testing

• One added README line; no code changes.

[mergify]

Merge Protections

🔴 1 of 1 protections blocking · waiting on 👀 reviews and 🤖 CI

	Protection	Waiting on
🔴	PR merge requirements	👀 reviews and 🤖 CI

🔴 PR merge requirements

Waiting for

• #approved-reviews-by>=1
• check-success=full-suite-passed

This rule is failing.

• #approved-reviews-by>=1
• check-success=full-suite-passed
• check-success=fastcheck-passed
• check-success~=pre-commit
• title~=(?i)^\[(feat|feature|bugfix|fix|refactor|perf|ci|doc|docs|misc|chore|kernel|new.?model|skill|skills|infra)\]

[greptile-apps]

Confidence Score: 4/5

[Low risk] Adds product announcement to the readme.

The PR should not merge until the DGX Spark V2 announcement matches a usable recipe or is qualified.

Findings

1.  Spark recipe runs V1 ▶
2.  Trim memory claim lacks conditions ▶

Fix with agent prompt

### Issue 1
README.md:12
The announcement says FastH3 V2 runs on one DGX Spark, but the linked cookbook identifies Spark as a V1-only runtime, and its single-Spark recipe loads the V1 checkpoint. A reader following that recipe will run V1 instead of the advertised V2. Please provide a V2 Spark recipe or qualify the announcement.

### Issue 2
README.md:12
The announcement says Trim runs in as little as 8 GB of GPU memory without identifying the hardware or inference settings needed to achieve that figure. The linked model is labeled NVFP4, and this repository’s NVFP4 runtime requires GPU capability sm100 or newer. Without those conditions or a reproducible Trim recipe, readers cannot tell whether their 8 GB GPU can run it.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

The PR adds a README news item announcing single-machine FastH3 V2 support and the experimental FastH3 Trim model.

• The DGX Spark portion conflicts with the linked V1-only recipe.
• The Trim memory claim needs hardware and inference conditions to be useful.

_{Reviews (1) · Last reviewed commit: "[docs] Announce FastH3 on consumer hardw..."}

[greptile-apps]

Spark recipe runs V1 The announcement says FastH3 V2 runs on one DGX Spark, but the linked cookbook identifies Spark as a V1-only runtime, and its single-Spark recipe loads the V1 checkpoint. A reader following that recipe will run V1 instead of the advertised V2. Please provide a V2 Spark recipe or qualify the announcement.

Prompt To Fix With AI

This is a comment left during a code review.
Path: README.md
Line: 12

Comment:
**Spark recipe runs V1** The announcement says FastH3 V2 runs on one DGX Spark, but the linked cookbook identifies Spark as a V1-only runtime, and its single-Spark recipe loads the V1 checkpoint. A reader following that recipe will run V1 instead of the advertised V2. Please provide a V2 Spark recipe or qualify the announcement.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Trim memory claim lacks conditions The announcement says Trim runs in as little as 8 GB of GPU memory without identifying the hardware or inference settings needed to achieve that figure. The linked model is labeled NVFP4, and this repository’s NVFP4 runtime requires GPU capability sm100 or newer. Without those conditions or a reproducible Trim recipe, readers cannot tell whether their 8 GB GPU can run it.

Prompt To Fix With AI

This is a comment left during a code review.
Path: README.md
Line: 12

Comment:
**Trim memory claim lacks conditions** The announcement says Trim runs in as little as 8 GB of GPU memory without identifying the hardware or inference settings needed to achieve that figure. The linked model is labeled NVFP4, and this repository’s NVFP4 runtime requires GPU capability sm100 or newer. Without those conditions or a reproducible Trim recipe, readers cannot tell whether their 8 GB GPU can run it.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!