Model datasheet · DeepSeek R1 distill
DeepSeek R1-0528 Qwen3-8B
R1's reasoning distilled into a Qwen3-8B body: frontier-class competition math in 8B — AIME'24 86.0 matches models 30× its size. A specialist: brilliant at math and logic, awkward at everything agentic.
- Vendor
- DeepSeek
- Architecture
- Dense · 8B
- Context
- 131,072 tokens
- License
- MIT
- Released
- 2025-05-29
- Vision
- No
→ GGUF (llama.cpp / LM Studio / Ollama) → MLX (Apple Silicon)
Architecture, from its own config.json
36 layers · 36 full attention
32 attention heads over 8 KV heads (GQA 4:1), head dim 128 · dense feed-forward · vocabulary 151,936
Where the memory goes: weights by band above, plus a KV cache of 1.1GB at 8K · 4.5GB at 32K · 18GB at 128K, plus the runtime's own working memory — dominated by the output projection over that vocabulary.
Inside one layer, from its own tensors
36 × full-attention layer
The green line down the right of each half is the residual: it leaves the spine above the norm, skips the operation entirely and rejoins at the +. Where attention is drawn open, q, k and v are scaled to their real output widths — a model that shares key and value heads across many query heads shows one wide box beside two narrow ones.
Around the stack: an embedding of 151,936 × 4,096, a final norm, and a separate output projection.
On the norm names: despite what it is called, post_attention_layernorm is drawn after the residual add here, which is where this family applies it — it normalises the input to the feed-forward, not the output of attention. Models that carry a separate pre_feedforward_layernorm use it the other way, and are drawn that way.
§1 Characteristics by quantization band
— Not yet rated · contributions welcome —
Q4–Q5 — 4–5 bit
5GB weights · e.g. Q4_K_M , mlx-4bit + KV: 1.1GB at 8K · 4.5GB at 32K · 18GB at 128K— Not yet rated · contributions welcome —
— Not yet rated · contributions welcome —
FP16 — full precision (fp16/bf16)
16GB weights · e.g. bf16 + KV: 1.1GB at 8K · 4.5GB at 32K · 18GB at 128Kdata checked Aug 2026
- vendor-model-card — AIME 2024 = 86 @ bf16 · vendor-reported · 2026-08-30
AIME 2025: 76.3, HMMT Feb 2025: 61.5 — vs base Qwen3-8B's 76.0 AIME'24. GPQA-Diamond (61.1) actually drops slightly below base.
data checked Aug 2026
- vendor-model-card — LiveCodeBench 2408–2505 = 60.5 @ bf16 · vendor-reported · 2026-08-30
Benchmark-decent, but community reports real-world coding friction (see known issues). Absent from Aider (only full R1 appears at 71.4%).
§2 Known issues & what fixes them
Unreliable tool/function calling — broken or ignored tool schemas are a recurring report across runtimes. This is not an agent model.
Community-fixed chat templates (unsloth's GGUFs) improve it somewhat.
For agents, use base Qwen3-8B and keep this one for pure reasoning.
Evidence · 2 sources · community-consensus
- Tool-calling reliability discussion (other, 2026-08-30)
- vLLM issue: R1-0528 distill function calling (github, 2026-08-30)
Very long, redundant chains of thought — it can spend thousands of tokens re-deriving a simple answer, which is slow and floods small context windows.
temp 0.6 per the card; cap max output tokens; save it for problems that deserve the thinking.
Evidence · 2 sources · community-consensus
- Study of reasoning-model overthinking (paper, 2026-08-30)
- Analysis of DeepSeek chain-of-thought verbosity (blog, 2026-08-30)