Abstract#
We describe Flare-1, the architecture behind North ML's Ink-1 (~725M
parameters), Fiber-1, and Fiber-1 Pro (~1.63B parameters). Flare-1 couples a
Qwen3.5-derived gated DeltaNet/GQA hybrid backbone ("Starlight") with a second,
smaller transformer trunk ("Fiber") that shares the backbone's embedding table
and reads its hidden states through a learned bridge. Two additions give the
merged system its character: AgentRegisters, small persistent memory slots
attached to the backbone's full-attention layers whose state survives across
inference calls, and DeltaMemory, delta-rule associative memory layers
inside the Fiber trunk. A linear router head on the pooled embedding gates a
retrieval-and-tool pipeline, and a state-management layer ("Era") provides
lossless snapshots, rollback, and prefix checkpoints for speculative decoding
and long-context work. All components ship in a single model.safetensors
artifact, 8-bit quantized, loadable on commodity hardware (a 16 GB Mac Mini).
We are deliberately honest about results: Flare-1 models are strong for their size, not frontier models. Fiber-1 scores 73.5% on ARC-Challenge and 37.2% on HumanEval; a 31-case held-out reliability suite with fictional entities and offline retrieval fixtures yields 19/20 (Fiber-1 Pro), 17/20 (Fiber-1), and 17/20 (Ink-1). A 1M-parameter controlled comparison of the Flare mechanism set versus a param-matched GPT baseline showed no significant advantage — we publish this negative result rather than claim unproven superiority.
1. Introduction#
Small open models are fluent but unreliable: stale knowledge, hallucinated facts, no tools. The premise of Flare-1 is that much of this unreliability is systemic, not parametric — a small model wrapped in correct routing, retrieval, and verification can answer grounded questions far above its raw weight class, while remaining cheap enough to run on a laptop.
Flare-1 is the merger of two prior North ML lines:
- Starlight — a Qwen3.5-style hybrid backbone (gated DeltaNet linear attention + periodic full GQA attention) distilled and light-pruned from a larger teacher.
- Fiber — a 158M-parameter trunk originally built as a router/draft model, extended with delta-rule associative memory.
The architecture goal: keep the backbone's trained behavior intact (all additions are zero-init or side-path) while adding persistence (registers), associative recall (DeltaMemory), and decision-making (router) as separable, measurable components.
2. Architecture#
2.1 Starlight backbone#
| Ink-1 | Fiber-1 / Fiber-1 Pro | |
|---|---|---|
| Layers | 24 | 24 |
| Hidden size | 1024 | 2048 |
| Pattern | 3× linear_attention + 1× full_attention (interval 4) | same |
| Attention | GQA, 8 query / 2 KV heads, head_dim 256 | same |
| MLP intermediate | 3584 | SwiGLU, scaled |
| Vocabulary | 65,536 | 65,536 |
| Parameters | 565,222,208 | 1,413,111,616 |
Linear-attention layers are gated DeltaNet blocks: per-head recurrent state
S ∈ R^{d×d} updated by the delta rule, with a learned per-step decay/gate.
Every fourth layer is full GQA attention with RoPE — this hybridization keeps
per-token cost close to linear attention while retaining periodic exact
retrieval over the sequence.
2.2 AgentRegisters#
After each full-attention layer, a small set of register slots is updated:
mem ← tanh(W_r · concat(h, mem_prev))
out ← h + W_o · mem (W_o zero-initialized)Because W_o starts at zero, registers are functionally invisible to the
trained backbone at initialization and only contribute once trained.
Crucially, register state is persisted across calls (not just across
tokens within a sequence): it acts as a session-level recurrent memory that
survives generation boundaries, snapshots, and rollbacks. Register state is a
first-class citizen of the runtime state — it is saved, restored, and
prefix-cached alongside KV and DeltaNet states.
Parameter cost is small: 295,008 (Ink-1) / 589,920 (Fiber-1).
2.3 Fiber trunk and DeltaMemory#
The Fiber trunk is a complete 24-layer transformer (hidden 768, 12 heads,
intermediate 2048, 158,143,524 params) that shares the backbone's 65,536-token
embedding table. Three of its layers are replaced by DeltaMemory layers:
delta-rule associative memory maintaining an S[H,D,D] matrix state read and
written each step — the same update family as DeltaNet, but dedicated to
key→value association rather than sequence modeling.
Fiber serves three roles:
- Router — pooled output feeds the search/think head (§2.4).
- Draft model — cheap candidate tokens for speculative decoding against Starlight verification.
- Associative memory — DeltaMemory state persists across calls like the registers.
2.4 Bridge and router head#
A learned linear bridge (1,572,864 params on Ink-1; 3,145,728 on Fiber-1)
projects Starlight hidden states into Fiber's width. The router head is a
1,538-parameter linear probe over the pooled representation emitting
p_search and p_think; thresholds (p_search ≈ 0.94) gate the retrieval
pipeline. The router is deliberately dumb — contextual rewriting upstream
turns follow-ups into standalone questions before routing, and raw-message
overrides keep small talk from ever triggering search.
2.5 Era (state management — not attention)#
Era is a lossless state manager, not a new attention mechanism. It snapshots and restores the complete runtime state — KV cache, DeltaNet recurrent states, AgentRegister memory, DeltaMemory matrices, and Fiber state — enabling:
- Rollback: rewind generation to any prior checkpoint.
- Prefix checkpoints: resume from a shared prompt prefix without re-prefill (disabled by default pending parity verification).
- Speculative decoding: Fiber drafts, Starlight verifies, mismatches roll back exactly.
2.6 Merged artifact#
Each tier is one model.safetensors (~755 MB / ~1.76 GB) containing all
1,368 tensors plus JSON metadata: starlight.config, fiber.config, router
thresholds, search config, generation defaults, prompt templates, chat
template, and tokenizer. Weight matrices use block-quantized 8-bit storage
([64,8] scheme, packed as U32); support tensors remain BF16/F32. A GGUF
container build is published for tooling that prefers the format (identical
tensors; not llama.cpp-compatible).
3. Training#
- Starlight (Fiber-1/pro): ~1.63B backbone produced by 95.8% light pruning + distillation from a larger Qwen3.5-style teacher.
- Ink-1: 565M Starlight-small variant, same recipe.
- Fiber-1 Pro: Fiber-1 weights + online GRPO RL plus a neural fact-verification pass that checks emitted claims against retrieved sources and revises unsupported ones.
- Fiber trunk and router are trained/fit as a draft-and-route auxiliary; the bridge is learned on embedding-paired data.
4. Retrieval and agentic pipeline#
At inference the Kestrel runtime runs: last-user-message → contextual rewrite (history-aware standalone question) → raw-message chat overrides → router → tools (weather/geocoding/time/math/Python) → retrieval (model-written query → Wikipedia + live web → BM25 + embedding-similarity rerank with recency and disambiguation weighting) → grounded generation with source citations → optional verification pass (Pro) → streamed stages to the client (routing/searching/reading/thinking/writing/verifying).
5. Evaluation#
5.1 Standard benchmarks (measured)#
| Benchmark | Ink-1 (725M) | Fiber-1 (1.63B) |
|---|---|---|
| ARC-Challenge (1165q) | 63.6% | 73.5% |
| MMLU (280q subset) | 39.3% | 48.9% |
| HumanEval (164) | 18.3% | 37.2% |
| GSM8K (300q) | 40.3% | unstable across runs |
5.2 Held-out reliability suite#
31 cases (chat, identity, grounding, follow-up, pushback, reasoning, code) with fictional entities and offline retrieval fixtures, judged semantically:
| Baseline runtime | Candidate runtime | |
|---|---|---|
| Ink-1 | 15/20 | 17/20 |
| Fiber-1 | 14/20 | 17/20 |
| Fiber-1 Pro | 15/20 | 19/20 |
The candidate runtime (contextual rewrite + grounded writer + raw-message routing) fixed follow-up handling (1/3 → 3/3) and Pro pushback (0/2 → 2/2) with no regressions.
5.3 Flare-1M controlled experiment (negative result)#
A 965,908-param Flare-mini (DeltaNet + GQA + registers + DeltaMemory + cross-chunk state) vs a 981,312-param GPT baseline, same data/steps:
| Metric | Flare-1M | GPT baseline |
|---|---|---|
| Held-out CE | 0.0004 | 0.0003 |
| Passkey @ depth .1/.5/.9 | 0 / .083 / .083 | 0 / 0 / .083 |
| Associative recall | .146 | .104 |
| Train time | 576 s | 36 s |
No significant advantage; the passkey task (2,600 tokens) exceeded the 384-token training window for both models, and near-zero held-out CE indicates corpus duplication. The architecture additions are not yet proven to help at small scale — a longer-context training run is needed to discriminate.
6. Limitations#
- English only; hallucination remains possible, especially under the small Ink-1 writer which can blend corrected and stale claims.
- The router is a linear probe — coverage errors route wrong.
- GSM8K scores are unstable across runs; treat as noise.
- Era prefix caching is disabled by default pending parity tests.
- No evidence the merged system beats frontier models on general capability; the pitch is capability per dollar and grounded verifiability.
7. Related work#
Gated linear attention (DeltaNet/DeltaRule family, Qwen3.5 hybrid backbone); mixture of recurrent + full-attention hybrids (Jamba, Zamba); retrieval- augmented generation; speculative decoding; differentiable memory slots (memory tokens, register tokens in ViT-LM hybrids).
8. Availability#
Weights: North-ML1/Ink-1, North-ML1/Fiber, North-ML1/Fiber-pro
(safetensors), North-ML1/{Ink-1,Fiber-1,Fiber-1-Pro}-GGUF (container),
North-ML1/Flare-Preview (architecture + POC code/results). Runtime: Kestrel
(MLX). All artifacts Apache-2.0.