Adil Eco V1.1 · Compact variant · Research release candidate
The compact Adil model, measured honestly.
Small enough for everyday CPUs. Not released yet.
Adil Eco is the compact variant of the Adil model family, intended for local inference on everyday CPU hardware. Its current candidate is a research release candidate: the release decision is no-go, and there is no download, package or commercial offer.
Research release candidate Not yet available
Release status
Release decision: no-go.
The current Eco candidate failed its sealed acceptance gate. The locked gate required a difference of at least 5 percentage points over the base model; the measured difference was 4.16 percentage points, so the gate failed.
A fresh sealed acceptance evaluation has been prepared but has not been run. Until a candidate passes it, and until licensing and owner approval are complete, Adil Eco will not be released.
Engineering benchmark
Compared with other compact models.
An internally conducted comparative engineering benchmark of the Adil Eco V1.1 candidate covers knowledge, maths, coding, instruction following, structured JSON, CPU memory and context latency. It is not independent third-party certification, and on most quality measures another compact model, LFM2-700M, scores higher. Mathematical reasoning regressed against Qwen3-0.6B Base in the same benchmark.
What Eco is
A compact model for bounded local tasks.
Adil Eco is adapted from an existing third-party base model by fine-tuning; it is not a foundation model trained from scratch.
It is intended for short text tasks inside applications, such as rewriting, extraction to JSON, classification and bounded assistance, with the host application keeping control of every action.
- Chat, streaming and JSON output worked in SDK qualification on development builds.
- The model emits tool calls, but tool calling through the SDK is under repair and not yet available.
- Adil Eco V1.1 will not support tool_choice set to required; such a request will return an explicit capability error.
- Thinking-style reasoning output is switched off in the developer preview.
Files and memory
File size is not memory use.
- Primary file: 4-bit (Q4_K_M) GGUF file of 484,219,776 bytes, about 484 MB.
- Quality-tier file: Q5_K_M, about 551 MB.
- The weights in the file are memory-mapped (462 MiB); the runtime also reserves a context cache of about 112 KiB per token, so memory use grows with the context length.
- Process footprint measured on an Apple M1 test host: about 867 MB at a 4K context and about 1,331 MB at the default 8K context.
- Process footprint measured on the same Apple M1 at the experimental 16K context: about 2,253 MB.
Measurement limits: Measured on one Apple M1 (8 GB) only. These figures are not minimum system requirements, and performance on x86 laptops such as the ThinkPad T480 is unverified.
Context length
8K by default. 16K is experimental.
Adil Eco is intended to use an 8K-token context by default.
A 16K context is experimental and limited: it worked functionally in tests, but with about 128 seconds to the first token and about 5 tokens per second on the Apple M1 test host.
Model card
Model card.
- Model
- Adil Eco V1.1 candidate: a research release candidate, not released.
- Identifiers
- In the developer preview's local API, the planned model identifier is adil-eco-v1.1, with the alias adil-eco and the display name Adil Eco V1.1.
- Base model
- Derived by fine-tuning from a third-party open-weight base model.
- Files
- A 4-bit (Q4_K_M) GGUF file, with a Q5_K_M quality-tier file.
- Context
- 8K tokens by default; 16K is experimental.
- Intended use
- Short, bounded text tasks inside applications, with the host application in control of every action.
- Out of scope
- Authorising payments, access or safety-critical actions, and unchecked factual or numerical answers.
- Evaluation
- The sealed acceptance gate failed, and the release acceptance has not been run.
- Licence and attribution
- The licence and the attribution required for the base model will be published with any release. Rights relating to the licence, the teacher data used in fine-tuning and third-party code headers bundled with the runtime are pending legal review.
FAQ
- Can I download Adil Eco?
- No. The release decision is no-go, Adil Eco is not available, and no download or package has been published.
- Does Eco run tools?
- No model runs tools by itself. The model emits tool calls, but tool calling through the SDK is under repair and not yet available; any tool would be run by your application, never by the model or the gateway.
- Will it run on my laptop?
- Only an Apple M1 has been measured. Performance on older x86 laptops, including the ThinkPad T480, is unverified.