emkey1

Qwen3.6-35B-A3B-Aether

README

License: apache-2.0

Results

All scores use no guide in the prompt, temperature 0, seed 42, one sample per task, and up to two repair rounds in which the model sees the compiler's diagnostic and the observed output.

"First attempt" counts tasks solved by the first program; "after repair" counts tasks solved within the repairs.
Graded with the Aether compiler at language version 2026-10-21-8 (1.0-beta4dev).
- Every earlier model replayed on it scores exactly as on 2026-10-20-12.
Columns: first attempt (after repair).

Table with columns: suite, tasks, untrained base, no guide, this model, no guide, untrained base + 14K-token guide
suite	tasks	untrained base, no guide	this model, no guide	untrained base + 14K-token guide
core	35	0 (0)	33 (35)	33 (35)
hard: TOON data processing	14	0 (0)	13 (14)	13 (14)
hard: other	5	0 (0)	2 (4)	1 (3)
algorithms	19	0 (0)	17 (17)	13 (17)
main total	73	0 (0)	65 (70), 89%	60 (69)
held-out: surface	15	0 (0)	10 (14)	12 (14)
held-out: algorithmic	14	0 (0)	10 (11)	5 (8)
held-out: specification	12	0 (0)	9 (11)	5 (7)
held-out: TOON / memory streams	12	0 (0)	10 (11)	7 (9)
held-out total	53	0 (0)	39 (47), 74%	29 (38)

With no guide, it beats the base model reading the 14K-token guide on both measures: 65 against 60 on the main suites, 39 against 29 on the held-out suites.
Held-out suites: these were never used to write training data or to mine failures. They measure how far the gains carry beyond the tasks the data was developed against.
- They were used once to choose between two finished checkpoints (next section), so the held-out score carries a small selection bias.
Against the dense 27B sibling (same data, graded the same way): it trails on the main suites (65 against 71) and matches it on the held-out suites (39 against 37, within run-to-run noise).
The untrained-base columns were measured with an 8-bit (int8) copy of the base model. On a smoke test, int8 and bf16 scoring differed by one task.

Choosing the checkpoint

Eval loss on a held-out slice of the training data was the wrong guide for this model.

With early stopping on eval loss, training stopped at epoch 0.76 and scored 62 (main) and 30 (held-out).
Training on for three full epochs raised eval loss (the model memorises its training set), yet it lifted the held-out score to 39.

Four finished models were compared, first attempt, main / held-out:

Table with columns: main (73), held-out (53)
	main (73)	held-out (53)
early-stopped (epoch 0.76)	62	30
epoch 2 of the 3-epoch run	68	33
a run scheduled for exactly 2 epochs	67	33
epoch 3, this model	65	39

A rule fixed before the last run would have picked the 2-epoch run, on a two-task first-attempt gap on the main suites; after repair the two tie at 70. This model was chosen instead for its held-out lead of six tasks, eight after repair. That decision used the held-out suites, which is the reason for the selection-bias note above.

Python, same model, same tasks

Table with columns: suite, tasks, untrained base, this model
suite	tasks	untrained base	this model
core	35	30 (32)	6 (18)
hard: TOON data processing	14	14 (14)	5 (11)
hard: other	5	5 (5)	0 (2)
algorithms	19	19 (19)	17 (19)
held-out: surface	15	15 (15)	0 (6)
held-out: algorithmic	14	14 (14)	10 (12)
held-out: specification	12	11 (12)	6 (9)
held-out: TOON / memory streams	12	9 (11)	0 (5)
total	126	117 (122)	44 (82)

Most failed Python first attempts are Aether syntax in a Python answer (68 of 82): fn main, let x: Int, fx {.
This is the forgetting that comes with specialising a model on one language, and Aether looks a lot like Python. Three epochs of Aether deepen it: the early-stopped run kept 78.
Both columns were measured with int8 copies on the same machine.

How to use

The model was trained with thinking off and with no system message. Serve it the same way:

vLLM: pass chat_template_kwargs={"enable_thinking": false} with each request.
Ollama (OpenAI-compatible /v1, merged model): send "reasoning_effort": "none".
- A top-level think: false is ignored on that endpoint, and the model will reason first.
- Set the context in a Modelfile (PARAMETER num_ctx 32768); /v1 ignores per-request context options.
- Importing the merged safetensors with ollama create -q int8 gives a 39 GB model, tested for this card.

Loading the merged model

text

vllm serve emkey1/Qwen3.6-35B-A3B-Aether --max-model-len 32768

Send "chat_template_kwargs": {"enable_thinking": false} with each request (thinking off, as trained). This is how the Aether scores above were measured (bf16).

Prompt format used in training and evaluation (a single user message):

text

You are writing Aether code. Do not assume you have an Aether reference guide for this task.Use only what you already know about the language and infer cautiously. Write exactly one complete Aether program that solves the task below. Requirements:- Return only raw Aether source code....Task:<task statement> Expected stdout:<exact expected output> Aether source:

Shorter natural-language requests ("Write an Aether program that ...") also work; part of the training data is in that form. Keep the Aether compiler in the loop: feeding its diagnostic back fixes most first-attempt misses.

Training

Method: LoRA r=32, alpha=64, dropout 0.
- LoRA covers the 256 routed experts (their fused gate_up_proj / down_proj weights), the shared expert, and the projections of the 10 full-attention layers: 1.86B trainable parameters, 5% of the model.
- The router is not trained.
- The 30 linear-attention layers carry no LoRA.
Loading. The base model was loaded in Unsloth's 4-bit mode. bitsandbytes cannot quantize the fused expert weights, so the experts, most of the model, stay in bf16; attention and the shared expert are 4-bit.
Optimisation.
- lr 1e-4, linear decay over 3 epochs; batch 1 × 8 gradient accumulation; seed 3407.
- Loss on responses only (prompt tokens masked). No system message on any record.
- No early stopping: the last step of epoch 3 is released (see "Choosing the checkpoint").
- Unsloth + TRL + PEFT on one GPU host, about 11 hours.
Data: the same 1,981 instruction records (1,761 Aether, 220 Python) and 87 repair records as the dense sibling, on Aether 1.0-beta. Every Aether program compiles and prints its stated output with the compiler, checked on every build. The records are:
- a curated corpus of Aether programs and natural-language task/solution pairs;
- repair drills: a broken program, the compiler's diagnostic, the fix;
- targeted packs for the idioms earlier runs got wrong (file reads inside fx, JSON/TOON, memory streams, modules, tuple contracts, array value semantics). They were written from failure classes on the main suites, never from task text;
- a standard-library pack covering every builtin family in the official list, and seed pairs for maps;
- distilled pairs:
  - task statements and independent Python solutions written by MiMo-V2.6-Flash-RL (MIT);
  - Aether solutions written with the guide, by an untrained base model or, for the tasks it could not solve, by MiMo;
  - kept only where the Aether program prints exactly what the Python program prints.
- Python pairs (MiMo's programs for the same tasks) to limit the drift away from Python.
Decontamination:
- Every training example is checked against every benchmark task: an identical expected output is dropped, and 5-gram similarity to any reference solution must stay below 0.5.
- The held-out suites are excluded from data authoring and from failure mining.

The language, its compiler, the guides, the benchmark tasks, the training data and every scoring report are at emkey1/aether (bench/, training/; this model is run qwen36-35b-a3b-e3).

Limitations

Aether only. Python, and probably other languages, are badly degraded; see above.
Language version. The model knows Aether 1.0-beta (2026-10-20). Aether 1.0 freezes the language: valid programs keep compiling with the same output, so later releases should add to what it knows rather than break it. Until the final 1.0, a break is still possible but needs an explicit, recorded decision.
Program size. It writes small, self-contained programs well. It has not been evaluated on multi-file projects or long programs.
First attempts still fail on some idioms that one repair round fixes. Use it with the compiler in the loop.

License and attribution

Apache-2.0, as the base model. Distilled training data was generated with MiMo-V2.6-Flash-RL (MIT, Xiaomi).

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

emkey1

Model Tree

Base

Qwen/Qwen3.6-35B-A3B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Results

All scores use no guide in the prompt, temperature 0, seed 42, one sample per task, and up to two repair rounds in which the model sees the compiler's diagnostic and the observed output.

"First attempt" counts tasks solved by the first program; "after repair" counts tasks solved within the repairs.
Graded with the Aether compiler at language version 2026-10-21-8 (1.0-beta4dev).
- Every earlier model replayed on it scores exactly as on 2026-10-20-12.
Columns: first attempt (after repair).

Table with columns: suite, tasks, untrained base, no guide, this model, no guide, untrained base + 14K-token guide
suite	tasks	untrained base, no guide	this model, no guide	untrained base + 14K-token guide
core	35	0 (0)	33 (35)	33 (35)
hard: TOON data processing	14	0 (0)	13 (14)	13 (14)
hard: other	5	0 (0)	2 (4)	1 (3)
algorithms	19	0 (0)	17 (17)	13 (17)
main total	73	0 (0)	65 (70), 89%	60 (69)
held-out: surface	15	0 (0)	10 (14)	12 (14)
held-out: algorithmic	14	0 (0)	10 (11)	5 (8)
held-out: specification	12	0 (0)	9 (11)	5 (7)
held-out: TOON / memory streams	12	0 (0)	10 (11)	7 (9)
held-out total	53	0 (0)	39 (47), 74%	29 (38)

With no guide, it beats the base model reading the 14K-token guide on both measures: 65 against 60 on the main suites, 39 against 29 on the held-out suites.
Held-out suites: these were never used to write training data or to mine failures. They measure how far the gains carry beyond the tasks the data was developed against.
- They were used once to choose between two finished checkpoints (next section), so the held-out score carries a small selection bias.
Against the dense 27B sibling (same data, graded the same way): it trails on the main suites (65 against 71) and matches it on the held-out suites (39 against 37, within run-to-run noise).
The untrained-base columns were measured with an 8-bit (int8) copy of the base model. On a smoke test, int8 and bf16 scoring differed by one task.

Choosing the checkpoint

Eval loss on a held-out slice of the training data was the wrong guide for this model.

With early stopping on eval loss, training stopped at epoch 0.76 and scored 62 (main) and 30 (held-out).
Training on for three full epochs raised eval loss (the model memorises its training set), yet it lifted the held-out score to 39.

Four finished models were compared, first attempt, main / held-out:

Table with columns: main (73), held-out (53)
	main (73)	held-out (53)
early-stopped (epoch 0.76)	62	30
epoch 2 of the 3-epoch run	68	33
a run scheduled for exactly 2 epochs	67	33
epoch 3, this model	65	39

Python, same model, same tasks

Table with columns: suite, tasks, untrained base, this model
suite	tasks	untrained base	this model
core	35	30 (32)	6 (18)
hard: TOON data processing	14	14 (14)	5 (11)
hard: other	5	5 (5)	0 (2)
algorithms	19	19 (19)	17 (19)
held-out: surface	15	15 (15)	0 (6)
held-out: algorithmic	14	14 (14)	10 (12)
held-out: specification	12	11 (12)	6 (9)
held-out: TOON / memory streams	12	9 (11)	0 (5)
total	126	117 (122)	44 (82)

Most failed Python first attempts are Aether syntax in a Python answer (68 of 82): fn main, let x: Int, fx {.
This is the forgetting that comes with specialising a model on one language, and Aether looks a lot like Python. Three epochs of Aether deepen it: the early-stopped run kept 78.
Both columns were measured with int8 copies on the same machine.

How to use

The model was trained with thinking off and with no system message. Serve it the same way:

vLLM: pass chat_template_kwargs={"enable_thinking": false} with each request.
Ollama (OpenAI-compatible /v1, merged model): send "reasoning_effort": "none".
- A top-level think: false is ignored on that endpoint, and the model will reason first.
- Set the context in a Modelfile (PARAMETER num_ctx 32768); /v1 ignores per-request context options.
- Importing the merged safetensors with ollama create -q int8 gives a 39 GB model, tested for this card.

Loading the merged model

text

vllm serve emkey1/Qwen3.6-35B-A3B-Aether --max-model-len 32768

Send "chat_template_kwargs": {"enable_thinking": false} with each request (thinking off, as trained). This is how the Aether scores above were measured (bf16).

Prompt format used in training and evaluation (a single user message):

text

You are writing Aether code.Do not assume you have an Aether reference guide for this task.Use only what you already know about the language and infer cautiously.Write exactly one complete Aether program that solves the task below.Requirements:- Return only raw Aether source code....Task:<task statement>Expected stdout:<exact expected output>Aether source:

You are writing Aether code. Do not assume you have an Aether reference guide for this task.Use only what you already know about the language and infer cautiously. Write exactly one complete Aether program that solves the task below. Requirements:- Return only raw Aether source code....Task:<task statement> Expected stdout:<exact expected output> Aether source:

Training

Method: LoRA r=32, alpha=64, dropout 0.
- LoRA covers the 256 routed experts (their fused gate_up_proj / down_proj weights), the shared expert, and the projections of the 10 full-attention layers: 1.86B trainable parameters, 5% of the model.
- The router is not trained.
- The 30 linear-attention layers carry no LoRA.
Loading. The base model was loaded in Unsloth's 4-bit mode. bitsandbytes cannot quantize the fused expert weights, so the experts, most of the model, stay in bf16; attention and the shared expert are 4-bit.
Optimisation.
- lr 1e-4, linear decay over 3 epochs; batch 1 × 8 gradient accumulation; seed 3407.
- Loss on responses only (prompt tokens masked). No system message on any record.
- No early stopping: the last step of epoch 3 is released (see "Choosing the checkpoint").
- Unsloth + TRL + PEFT on one GPU host, about 11 hours.
Data: the same 1,981 instruction records (1,761 Aether, 220 Python) and 87 repair records as the dense sibling, on Aether 1.0-beta. Every Aether program compiles and prints its stated output with the compiler, checked on every build. The records are:
- a curated corpus of Aether programs and natural-language task/solution pairs;
- repair drills: a broken program, the compiler's diagnostic, the fix;
- targeted packs for the idioms earlier runs got wrong (file reads inside fx, JSON/TOON, memory streams, modules, tuple contracts, array value semantics). They were written from failure classes on the main suites, never from task text;
- a standard-library pack covering every builtin family in the official list, and seed pairs for maps;
- distilled pairs:
  - task statements and independent Python solutions written by MiMo-V2.6-Flash-RL (MIT);
  - Aether solutions written with the guide, by an untrained base model or, for the tasks it could not solve, by MiMo;
  - kept only where the Aether program prints exactly what the Python program prints.
- Python pairs (MiMo's programs for the same tasks) to limit the drift away from Python.
Decontamination:
- Every training example is checked against every benchmark task: an identical expected output is dropped, and 5-gram similarity to any reference solution must stay below 0.5.
- The held-out suites are excluded from data authoring and from failure mining.

The language, its compiler, the guides, the benchmark tasks, the training data and every scoring report are at emkey1/aether (bench/, training/; this model is run qwen36-35b-a3b-e3).

Limitations

Aether only. Python, and probably other languages, are badly degraded; see above.
Language version. The model knows Aether 1.0-beta (2026-10-20). Aether 1.0 freezes the language: valid programs keep compiling with the same output, so later releases should add to what it knows rather than break it. Until the final 1.0, a break is still possible but needs an explicit, recorded decision.
Program size. It writes small, self-contained programs well. It has not been evaluated on multi-file projects or long programs.
First attempts still fail on some idioms that one repair round fixes. Use it with the compiler in the loop.

License and attribution

Apache-2.0, as the base model. Distilled training data was generated with MiMo-V2.6-Flash-RL (MIT, Xiaomi).