agentenv · shouc · Jul 29, 2026 · Jul 29, 2026
diff --git a/README.md b/README.md
@@ -2,16 +2,11 @@
 
 **Yeto** fine-tunes language and diffusion models across cheap, geographically
 scattered GPU capacity — spot instances, mixed regions, mixed clouds, even
-mixed hardware families — via the [SkyPilot](https://skypilot.co) SDK. Its
-asynchronous synchronization is based on **Decoupled DiLoCo**
-([Douillard et al., arXiv 2604.21428](https://arxiv.org/abs/2604.21428)):
-a Rust syncer merges parameter fragments from independent learner islands
-(quorum + adaptive grace, token-weighted RDA, Nesterov outer step), so slow
-links and preempted islands never block training.
+mixed hardware families.
 
 ```
                  ┌──────────────────────────────┐
-                 │  syncer (Rust, hot path)     │
+                 │  syncer (hot path)           │
                  │  fragment ingest · RDA merge │
                  │  Nesterov outer step · bcast │
                  └──────┬───────┬───────┬───────┘
@@ -60,9 +55,16 @@ yeto status | logs <run> | down <run>   # runs detach; Ctrl-C never kills them
 
 ## Architecture
 
+Fleets launch via the [SkyPilot](https://skypilot.co) SDK; asynchronous
+synchronization is based on **Decoupled DiLoCo**
+([Douillard et al., arXiv 2604.21428](https://arxiv.org/abs/2604.21428)):
+the syncer merges parameter fragments from independent learner islands
+(quorum + adaptive grace, token-weighted RDA, Nesterov outer step), so slow
+links and preempted islands never block training.
+
 One protocol, one syncer, four learner backends. Every backend speaks the
-same fragment protocol to the Rust syncer and runs the same DiLoCo step
-boundary — the pull/merge/α-blend/push loop lives in one shared module
+same fragment protocol and runs the same DiLoCo step boundary — the
+pull/merge/α-blend/push loop lives in one shared module
 (`yeto/diloco_sync.py`), so protocol changes land once and apply everywhere.
 
 | backend | selector | scope | validation |
@@ -78,53 +80,6 @@ terminal contract — budget cutoff, authoritative final cut, checkpoint
 marking — is shared by all backends (`finalization.py`,
 `budget_finalization.py`, `final_marker.py`).
 
-## Production source safety
-
-Remote model and dataset branches or tags are resolved to immutable Hugging
-Face commits before any learner is provisioned. Config, tokenizer, model,
-dataset, prefetch, export, and sampling paths reuse those commits. Executable
-model code is disabled by default; enabling `--trust-remote-code` is an
-explicit trust decision and still uses the pinned commit. Saved artifacts
-include `yeto_provenance.json` (or embedded diffusion provenance) with
-resolved source identities and a Yeto source-tree digest.
-
-Legacy pickled losses are disabled by default because deserialization executes
-code. Existing custom/callable transport requires
-`--allow-unsafe-pickled-loss` and is SHA-256 attested before every learner
-loads it. See [docs/PROVENANCE.md](docs/PROVENANCE.md) for revision flags,
-local-path behavior, custom diffusion-loader requirements, safe tensor
-formats, and migration guidance.
-
-## Diffusion
-
-Yeto's diffusion learner (`--model-kind diffusion`) keeps the LM-style
-contract: aliases are only repo-id shortcuts and raw Hugging Face ids pass
-through unchanged. Models that Diffusers can load use the generic path first —
-Yeto loads the pipeline, discovers the trainable denoiser, attaches LoRA,
-builds conditioning from data rows, and reuses the normal DiLoCo sync
-machinery. Model-family behavior that the generic path cannot infer lives in
-behavior adapters under `yeto/diffusion/adapters/` (in-tree: PixArt); use
-`--diffusion-adapter module:factory` for external models that need one.
-
-Recover the authoritative merged adapter from a syncer checkpoint with:
-
-```bash
-yeto-diffusion-export --checkpoint yeto-state.ckpt --model <id> --output-dir out/
-```
-
-Pass the same LoRA, fragment-pattern, and external-adapter flags used by the
-training run.
-
-- `--diffusion-seed <integer>`: reproducible initialization, data ordering,
-  timestep sampling, and noise (RNG streams only; does not force
-  deterministic CUDA kernels). Diffusion-only.
-- `--resize-mode center-crop` with `--height`/`--width`: preserve aspect
-  ratio by scaling to fill the target then center-cropping (default is
-  direct resize).
-
-The complete backend architecture, data contract, artifact flow, adapter
-boundary, and current limitations are in [docs/DIFFUSION.md](docs/DIFFUSION.md).
-
 ## Supported models
 
 Aliases are sugar over `yeto/models.py` (this table is generated by
@@ -133,17 +88,29 @@ is the frozen-base footprint an island's GPUs must jointly hold (bf16 base,
 LoRA) — add ~8 GB per GPU for activations/overhead, ×8 for full tuning;
 "(Hub)" means the size is resolved from safetensors metadata at plan time.
 
+Tested — a completed Yeto fine-tuning run on real hardware:
+
 | alias | Hugging Face id | min island VRAM (GB) |
 |---|---|---|
-| `gemma4` | `google/gemma-4-12B-it` | 66 |
 | `deepseek4flash` | `deepseek-ai/DeepSeek-V4-Flash` | 568 |
+| `qwen3-0.6b` | `Qwen/Qwen3-0.6B` | (Hub) |
+| `qwen36-27b` | `Qwen/Qwen3.6-27B` | 54 |
+| `kimi-k3` | `moonshotai/Kimi-K3` | (Hub) |
+| `glm52` | `zai-org/GLM-5.2` | 1488 |
+| `laguna-s-2.1` | `poolside/Laguna-S-2.1` | (Hub) |
+
+<details>
+<summary>All other supported aliases (untested)</summary>
+
+| alias | Hugging Face id | min island VRAM (GB) |
+|---|---|---|
+| `gemma4` | `google/gemma-4-12B-it` | 66 |
 | `qwen3-8b` | `Qwen/Qwen3-8B` | 17 |
 | `qwen35-4b` | `Qwen/Qwen3.5-4B` | 8 |
 | `qwen35-9b` | `Qwen/Qwen3.5-9B` | 18 |
 | `qwen35-9b-base` | `Qwen/Qwen3.5-9B-Base` | 18 |
 | `qwen35-35b-a3b` | `Qwen/Qwen3.5-35B-A3B-Base` | 70 |
 | `qwen35-397b-a17b` | `Qwen/Qwen3.5-397B-A17B` | 794 |
-| `qwen36-27b` | `Qwen/Qwen3.6-27B` | 54 |
 | `qwen36-35b-a3b` | `Qwen/Qwen3.6-35B-A3B` | 70 |
 | `llama32-1b` | `meta-llama/Llama-3.2-1B` | 3 |
 | `llama32-3b` | `meta-llama/Llama-3.2-3B` | 7 |
@@ -164,7 +131,6 @@ LoRA) — add ~8 GB per GPU for activations/overhead, ×8 for full tuning;
 | `nemotron3-ultra` | `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16` | 1100 |
 | `glm45-air` | `zai-org/GLM-4.5-Air` | 212 |
 | `glm46` | `zai-org/GLM-4.6` | 714 |
-| `glm52` | `zai-org/GLM-5.2` | 1488 |
 | `llama4-scout` | `meta-llama/Llama-4-Scout-17B-16E-Instruct` | 218 |
 | `llama4-maverick` | `meta-llama/Llama-4-Maverick-17B-128E-Instruct` | 800 |
 | `qwen3-coder-480b` | `Qwen/Qwen3-Coder-480B-A35B-Instruct` | 960 |
@@ -190,6 +156,8 @@ LoRA) — add ~8 GB per GPU for activations/overhead, ×8 for full tuning;
 | `deepseek31-base-bf16` | `unsloth/DeepSeek-V3.1-Base-BF16` | 1343 |
 | `kimi-k2-base-bf16` | `unsloth/Kimi-K2-Base-BF16` | 2060 |
 
+</details>
+
 ## Docs
 
 [docs/DESIGN.md](docs/DESIGN.md) — merge math, blending, adaptive grace,

diff --git a/scripts/gen_model_table.py b/scripts/gen_model_table.py
@@ -11,15 +11,16 @@
 import sys
 
 sys.path.insert(0, str(pathlib.Path(__file__).resolve().parents[1]))
-from yeto.models import MODEL_ALIASES, MODEL_WEIGHT_GB  # noqa: E402
+from yeto.models import MODEL_ALIASES, MODEL_WEIGHT_GB, TESTED_ALIASES  # noqa: E402
 
 SECTION_START = "## Supported models"
 SECTION_END = "## Docs"
 
 
-def table() -> str:
+def table(aliases) -> str:
     rows = ["| alias | Hugging Face id | min island VRAM (GB) |", "|---|---|---|"]
-    for alias, hf in MODEL_ALIASES.items():
+    for alias in aliases:
+        hf = MODEL_ALIASES[alias]
         gb = MODEL_WEIGHT_GB.get(alias)
         size = f"{gb:g}" if gb is not None else "(Hub)"
         rows.append(f"| `{alias}` | `{hf}` | {size} |")
@@ -35,7 +36,16 @@ def section() -> str:
 LoRA) — add ~8 GB per GPU for activations/overhead, ×8 for full tuning;
 "(Hub)" means the size is resolved from safetensors metadata at plan time.
 
-{table()}
+Tested — a completed Yeto fine-tuning run on real hardware:
+
+{table(a for a in MODEL_ALIASES if a in TESTED_ALIASES)}
+
+<details>
+<summary>All other supported aliases (untested)</summary>
+
+{table(a for a in MODEL_ALIASES if a not in TESTED_ALIASES)}
+
+</details>
 
 """
 

diff --git a/tests/test_models.py b/tests/test_models.py
@@ -42,7 +42,7 @@ def test_readme_table_matches_alias_table():
     import re
 
     readme = (pathlib.Path(__file__).resolve().parents[1] / "README.md").read_text()
-    table_aliases = set(re.findall(r"^\| `([a-z0-9\-]+)` \| `", readme, flags=re.M))
+    table_aliases = set(re.findall(r"^\| `([a-z0-9.\-]+)` \| `", readme, flags=re.M))
     assert table_aliases == set(MODEL_ALIASES), (
         "README model table is out of sync with yeto/models.py — regenerate it"
     )

diff --git a/yeto/models.py b/yeto/models.py
@@ -28,6 +28,7 @@
     "gemma4": "google/gemma-4-12B-it",
     "deepseek4flash": "deepseek-ai/DeepSeek-V4-Flash",
     # Qwen dense + MoE
+    "qwen3-0.6b": "Qwen/Qwen3-0.6B",
     "qwen3-8b": "Qwen/Qwen3-8B",
     "qwen35-4b": "Qwen/Qwen3.5-4B",
     "qwen35-9b": "Qwen/Qwen3.5-9B",
@@ -50,6 +51,7 @@
     "kimi-k2-thinking": "moonshotai/Kimi-K2-Thinking",
     "kimi-k25": "moonshotai/Kimi-K2.5",
     "kimi-k26": "moonshotai/Kimi-K2.6",
+    "kimi-k3": "moonshotai/Kimi-K3",  # size via Hub metadata
     # DeepSeek
     "deepseek31": "deepseek-ai/DeepSeek-V3.1",
     "deepseek-r1": "deepseek-ai/DeepSeek-R1",
@@ -71,6 +73,7 @@
     "minimax-m3": "MiniMaxAI/MiniMax-M3",  # size via Hub metadata
     "kimi-k27-code": "moonshotai/Kimi-K2.7-Code",  # size via Hub metadata
     "mistral-small3": "mistralai/Mistral-Small-3.2-24B-Instruct-2506",
+    "laguna-s-2.1": "poolside/Laguna-S-2.1",  # 118B-A8B agentic coding; size via Hub metadata
     # Ornith (DeepReinforce agentic-coding family, MIT)
     "ornith-9b": "deepreinforce-ai/Ornith-1.0-9B",
     "ornith-31b": "deepreinforce-ai/Ornith-1.0-31B",
@@ -105,6 +108,17 @@
     # fp8 repo yourself and pass the path via --model.
 }
 
+# Aliases with a completed Yeto fine-tuning run on real hardware. The
+# README surfaces these first; everything else is supported but untested.
+TESTED_ALIASES = {
+    "qwen36-27b",
+    "deepseek4flash",
+    "qwen3-0.6b",
+    "glm52",
+    "kimi-k3",
+    "laguna-s-2.1",
+}
+
 DIFFUSION_MODEL_ALIASES = {
     "flux": "black-forest-labs/FLUX.1-dev",
     "flux-schnell": "black-forest-labs/FLUX.1-schnell",