emkey1

Qwen3.8-27B-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.
Every number below is graded with the Aether compiler at language version 2026-10-20-12 (1.0-beta2).
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 (2)	35 (35)	33 (35)
hard: TOON data processing	14	0 (0)	14 (14)	13 (13)
hard: other	5	0 (0)	3 (4)	4 (4)
algorithms	19	0 (1)	19 (19)	17 (18)
main total	73	0 (3)	71 (72), 97%	67 (70)
held-out: surface	15	0 (0)	9 (12)	15 (15)
held-out: algorithmic	14	0 (0)	10 (13)	10 (13)
held-out: specification	12	0 (0)	10 (12)	12 (12)
held-out: TOON / memory streams	12	0 (0)	8 (9)	10 (12)
held-out total	53	0 (0)	37 (46), 70%	47 (52)

Main suites: with no guide, the model beats the base model reading a 14K-token guide on the first attempt (71 against 67), and solves all but one task after repair.
Held-out suites: these were never used to choose or write training data, so they measure how far the gains carry beyond the tasks the training data was developed against. There the guided base is still ahead.

Run-to-run variation: this revision is a single training run. Three seeds of the same recipe on the earlier data spread by one task (68–69 of 73 main, 39–40 of 53 held-out), so differences of a task or two between revisions are noise.

Changes from the first release

Graded on the same compiler (2026-10-20-12), the first release's archived programs score 62 of 73 main and 37 of 53 held-out on the first attempt. Six of its main-suite misses are a single rename: Aether 1.0 spells formatfloat as format_float, and the first release learned the old name. This revision:

writes the Aether 1.0-beta names throughout (snake_case builtins, try_ I/O variants);
was trained with maps and with 211 of the 213 builtins in the official list;
scores 71 (main) and 37 (held-out) on the first attempt on that compiler.

Python, same model, same tasks

Table with columns: suite, tasks, untrained base, with the Aether adapter
suite	tasks	untrained base	with the Aether adapter
core	35	29 (33)	14 (30)
hard: TOON data processing	14	10 (13)	6 (13)
hard: other	5	5 (5)	3 (5)
algorithms	19	19 (19)	16 (19)
held-out: surface	15	15 (15)	3 (13)
held-out: algorithmic	14	13 (14)	10 (14)
held-out: specification	12	12 (12)	11 (12)
held-out: TOON / memory streams	12	11 (12)	4 (10)
total	126	114 (123)	67 (116)

Almost every Python first attempt the adapter gets wrong is Aether syntax in a Python answer (fn main, let x: Int, fx {). Shown the Python error, it usually recovers. This is the familiar forgetting that comes with specialising a model on one language, and Aether looks a lot like Python. Mixing Python pairs into the training data held it off in short training runs but not at the length that gives the best Aether. This revision, trained on more Aether, leaks more on the first attempt than the first release did (67 against 85 of 126; both reach 116 after repair). Hence the advice: load the adapter for Aether only.

How to use

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

vLLM / transformers: pass chat_template_kwargs={"enable_thinking": false}. With vLLM, --enable-lora should serve the base model and the adapter side by side (not tested by us).
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.

Loading the merged model

text

vllm serve emkey1/Qwen3.8-27B-Aether --max-model-len 32768

Send "chat_template_kwargs": {"enable_thinking": false} with each request (thinking off, as trained).

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, on the attention and MLP projections.
- Base model loaded in 4-bit (bitsandbytes); lr 1e-4; batch 1 × 8 gradient accumulation; seed 3407.
- Loss on responses only (prompt tokens masked). No system message on any record.
- Early stopping on 60 held-out training records, evaluated every 10 steps, patience 3. It stopped at epoch 1.04; the best checkpoint, at epoch 0.92, is the one released.
- Unsloth + TRL + PEFT on one GPU host, about 2 hours.
Data: 1,981 instruction records (1,761 Aether, 220 Python) and 87 repair records, 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 (SQLite in memory, HTTP without a network, sockets over loopback, files, environment, tasks), 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 the 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;
  - all generation requests audited to come from MiMo.
- 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 the scoring reports are at emkey1/aether (bench/, training/).

Limitations

Aether only. Python, and probably other languages, are degraded with the adapter loaded; 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. Examples: an invented builtin name (toon_text_array, mstream_from_file), a wrong type in an assignment, or mutating an array parameter under Aether's value semantics. 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/Qwen3.8-27B

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.
Every number below is graded with the Aether compiler at language version 2026-10-20-12 (1.0-beta2).
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 (2)	35 (35)	33 (35)
hard: TOON data processing	14	0 (0)	14 (14)	13 (13)
hard: other	5	0 (0)	3 (4)	4 (4)
algorithms	19	0 (1)	19 (19)	17 (18)
main total	73	0 (3)	71 (72), 97%	67 (70)
held-out: surface	15	0 (0)	9 (12)	15 (15)
held-out: algorithmic	14	0 (0)	10 (13)	10 (13)
held-out: specification	12	0 (0)	10 (12)	12 (12)
held-out: TOON / memory streams	12	0 (0)	8 (9)	10 (12)
held-out total	53	0 (0)	37 (46), 70%	47 (52)

Main suites: with no guide, the model beats the base model reading a 14K-token guide on the first attempt (71 against 67), and solves all but one task after repair.
Held-out suites: these were never used to choose or write training data, so they measure how far the gains carry beyond the tasks the training data was developed against. There the guided base is still ahead.

Changes from the first release

writes the Aether 1.0-beta names throughout (snake_case builtins, try_ I/O variants);
was trained with maps and with 211 of the 213 builtins in the official list;
scores 71 (main) and 37 (held-out) on the first attempt on that compiler.

Python, same model, same tasks

Table with columns: suite, tasks, untrained base, with the Aether adapter
suite	tasks	untrained base	with the Aether adapter
core	35	29 (33)	14 (30)
hard: TOON data processing	14	10 (13)	6 (13)
hard: other	5	5 (5)	3 (5)
algorithms	19	19 (19)	16 (19)
held-out: surface	15	15 (15)	3 (13)
held-out: algorithmic	14	13 (14)	10 (14)
held-out: specification	12	12 (12)	11 (12)
held-out: TOON / memory streams	12	11 (12)	4 (10)
total	126	114 (123)	67 (116)

How to use

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

vLLM / transformers: pass chat_template_kwargs={"enable_thinking": false}. With vLLM, --enable-lora should serve the base model and the adapter side by side (not tested by us).
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.

Loading the merged model

text

vllm serve emkey1/Qwen3.8-27B-Aether --max-model-len 32768

Send "chat_template_kwargs": {"enable_thinking": false} with each request (thinking off, as trained).

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, on the attention and MLP projections.
- Base model loaded in 4-bit (bitsandbytes); lr 1e-4; batch 1 × 8 gradient accumulation; seed 3407.
- Loss on responses only (prompt tokens masked). No system message on any record.
- Early stopping on 60 held-out training records, evaluated every 10 steps, patience 3. It stopped at epoch 1.04; the best checkpoint, at epoch 0.92, is the one released.
- Unsloth + TRL + PEFT on one GPU host, about 2 hours.
Data: 1,981 instruction records (1,761 Aether, 220 Python) and 87 repair records, 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 (SQLite in memory, HTTP without a network, sockets over loopback, files, environment, tasks), 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 the 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;
  - all generation requests audited to come from MiMo.
- 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 the scoring reports are at emkey1/aether (bench/, training/).

Limitations

Aether only. Python, and probably other languages, are degraded with the adapter loaded; 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. Examples: an invented builtin name (toon_text_array, mstream_from_file), a wrong type in an assignment, or mutating an array parameter under Aether's value semantics. 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).