forever-ai/Apeiron-1.0

Apeiron 1.0 — The First Model in the Ape's Line by Forever AI

Apeiron 1.0 is a coding AI agent by Forever AI (formerly Klox AI).
The first model in the proprietary Ape's line. Knowledge cutoff — October 2026.
Trained from scratch as a full model — not an adapter, not a fine-tune on top of someone else's base model.

---

1. What It Is

Apeiron 1.0 is a developer tool, not a chatbot. It:

• writes code from scratch based on a spec (Python, JavaScript, SQL, Bash, etc.);
• fixes bugs — finds the root cause from a traceback and provides a fix;
• explains solutions and complexity (Big-O notation in plain language);
• designs architecture: endpoints, schemas, project structure;
• writes tests (pytest, etc.) and reviews others' code against a checklist;
• works as an agent: clarifies the task → provides code → suggests tests and next steps.

Works in Russian and English, responds in the language of the request.
This is a full model trained from scratch by the Forever AI team, with no use of pre-trained base weights and no LoRA.

2. Numbers

Parameter	Value
Version	1.0, first in the Ape's line
Creator	Forever AI (formerly Klox AI)
Type	Full decoder-only model, trained from scratch
Parameters	~4B
Training examples	200,000
Training steps	982 (1 epoch)
Final loss	0.15
Token accuracy	0.95
Format	GGUF (F16, Q8_0, Q4_K_M)
Knowledge cutoff	October 2026
Languages	RU + EN

3. Proprietary Dataset (200,000 examples)

The dataset was assembled by the Forever AI team specifically for this product:

• Identity (~900 examples): name, version, creator, line, knowledge cutoff, specialization — dozens of question formulations in two languages so the model introduces itself consistently in any context.
• Autumn 2026 Knowledge (~1,000 examples): the Claude 5.5 family — Opus 5.5 (released 09/22/2026), Sonnet 5.5 (09/28/2026), Haiku 5.5 (10/07/2026), Fable 5.1; dates, prices, 1M token context, pairwise comparisons, model selection by task.
• Coding Tasks (~198,000 examples): 34 families — algorithms, sorting, retry logic, FastAPI, SQL and JOINs, pytest, regex, file handling, Git, Docker, Bash, analysis of 10 common errors, review, complexity estimation, asyncio, decorators, pandas, debounce, data structures, binary search, CLI, logging, curl, cron. Each example is unique: numbers, tables, endpoints, and code styles are randomized.

All 200,000 rows were validated: valid JSON, system/user/assistant format, zero duplicates, zero empty responses.

4. Behavior and Character

System identity is baked into the weights, not glued on via a prompt:

• introduces itself as Apeiron 1.0 by Forever AI in any dialogue;
• maintains the coding agent style: to the point, with ready code;
• does not invent facts about unfamiliar releases — honestly states where the knowledge cutoff ends;
• remembers that it is the first model in the Ape's line.

5. Format and Launch

This repository provides GGUF quantizations of Apeiron 1.0 (built on top of our full-precision FP16 weights at forever-ai/apeiron-1.0-merged). Available quantizations:

• F16 — maximum fidelity, largest file
• Q8_0 — near-lossless, good balance
• Q4_K_M — small and fast, recommended for most local setups

Runs out of the box with:

• llama.cpp
• Ollama
• LM Studio

For the full-precision safetensors weights and the Transformers / vLLM / SGLang stack, use the full-precision repository.

6. Benchmarks

Our models are evaluated and compared with leading on-device models of comparable size across a diverse set of tasks, including agentic tasks, code, mathematics, general tasks, and knowledge tasks.

Benchmark	Apeiron 1.0	Spark‑X2.5‑4B	Spark‑X2.5‑1.7B	Qwen3.5‑9B	Qwen3.5‑4B	Qwen3.5‑2B	Gemma4‑12B	Gemma4‑E4B	Gemma4‑E2B
Agent
BFCL‑V4	75.0	65.1	46.9	66.1*	50.3*	43.6*	37.4	36.9	30.2
τ²‑bench	91.0	75.1	65.3	79.1*	79.9*	48.8*	69.0*	42.2*	24.5*
τ³‑bench	39.8	30.4	20.1	9.3	6.7	4.1	13.3	10.1	8.8
MCP‑Atlas	63.7	54.6	23.4	47.4*	40.8*	14.8	30.5*	15.0*	12.6
MCP‑Mark	16.3	14.2	2.3	13.4	12.5	–	–	–	–
Workspace Bench	35.9	31.2	18.9	25.5	21.3	7.7	–	–	–
VitaBench2.0	29.0	25.2	8.3	15.6	18.2	5.2	12.4	4.8	4.4
BrowseComp	45.2	40.9	29.7	8.3	14.3	3.1	10.0	8.3	3.7
Code
SWE‑Bench Pro	51.1	44.4	10.4	33.8*	29.4*	1.9	21.9*	4.0*	–
SWE‑Bench Verified	61.1	41.6	28.3	53.1*	38.8*	6.8	44.2*	14.0*	–
SWE‑Bench Multilingual	61.3	53.3	23.3	43.3	27.7	5.0	32.5*	–	–
SciCode	42.1	34.7	18.2	32.7*	24.0	6.0	39.8	27.5	20.5
Math
Gaokao 2026	155.8	133.4	114.8	135.5	130.3	94.0	130.6	102.4	81.8
AIME 2026	99.6	90.7	69.4	88.2	83.0	30.8	82.1*	42.5*	37.5*
HMMT Feb 2026	97.1	81.2	48.4	70.8	69.7	21.5	65.6	34.2	20.5
IMO‑AnswerBench	85.3	74.2	45.4	69.8	68.5	–	57.2	26.9	22.6
General & Knowledge
IFEval	99.5	93.0	89.5	91.5*	89.8*	78.6*	94.8	45.3	34.8
IFBench	86.2	75.0	66.3	64.5	59.2	41.3*	73.5*	44.0*	22.7
AA‑LCR	72.5	56.3	24.3	63.0*	57.0*	25.6*	55.3*	34.7	18.3
HLE	16.1	12.3	6.3	14.3	8.6	2.1	13.1	3.9	2.5
GPQA	88.8	67.4	43.8	77.2	67.2	44.6	72.8	54.5	43.8

7. Use Cases

• Personal coding assistant in IDE and chat.
• Crash analysis and pre-commit review.
• Test and documentation generation for code.
• Education: explaining algorithms and debugging for beginners.
• Sub-agent for routine tasks in combination with larger models.

8. Limitations (Honest)

• A 4B-class model: verify code before running, especially anything involving security, money, or production.
• Knowledge is frozen at October 2026 — for newer releases, consult official documentation.
• Works best on short, focused tasks; very long contexts are not its profile.
• Quantized builds (Q8_0, Q4_K_M) trade a small amount of quality for size and speed.

9. Ape's Line Roadmap

• Apeiron 1.x: expanded coding dataset families, further inference optimizations.
• Apeiron 2.0: improved coding capabilities, fresher knowledge, more languages.
• The line evolves in iterations: each version is a measurable step in loss, accuracy, and task coverage — not just a number change.

10. FAQ

Is this a chatbot?
No. It is a coding agent: a response without a ready solution and next steps is the exception, not the norm.

How is it different from open base models?
Out-of-the-box behavior: name, agent character, knowledge cutoff set to autumn 2026, and 200,000 examples focused on development. Plus, the model is trained from scratch for this task.

What was it tested on?
Control questions like "Who are you?", knowledge sets on 5.5 releases, and coding tasks from its own dataset — identity answers are stable, code executes.

11. Creator

Forever AI (before rebranding — Klox AI). The product, dataset, and training cycle are our work. Apeiron 1.0 opens the Ape's line — more to come.

12. Responsibility

The model is a tool. Verify generated code, do not run blindly what you do not understand, and comply with the laws of your jurisdiction.