Particle-1
Particle-1 is Particle AI's first model (released October 28, 2026): a rank-64 LoRA post-train of DeepSeek V4.1 Flash (~1.25% of parameters) that teaches the model to reach correct agentic-coding answers with 28.4%–52.2% fewer output tokens at roughly accuracy parity, priced at $0.05/$0.50 per 1M tokens with a 1,048,576-token context and 131,072-token output budget.
Empirical Evaluation Results
Architectural Profile & Capabilities
Particle-1 (OpenRouter: particle-ai/particle-1) is a token-efficiency-optimized derivative of DeepSeek V4.1 Flash rather than a from-scratch foundation model. Particle AI's approach is a narrow, deliberate post-train: a rank-64 LoRA adapter covering roughly 1.25% of the base model's parameters, trained with RL-style objectives to shorten reasoning chains and tool-call verbosity on agentic software-engineering tasks. The backbone's weights, tokenizer, chat template, and sparse-attention stack are inherited from DeepSeek V4.1 Flash, an MoE transformer reported at 552B backbone parameters with 8B active parameters per prompt token and 16B per decoded token, DeepSeek Sparse Attention (DSA) indexing, and Engram conditional memory. Because the intervention is adapter-level, Particle-1 inherits all of that architecture plus the base model's 1,048,576-token context window and 131,072-token maximum completion budget; Particle AI has not published a first-party model card, parameter count, or standalone technical report, so architecture statements should be attributed to the inherited base rather than to independent specification of Particle-1. The published evaluation is tightly scoped to four agentic coding suites — DeepSWE 1.1, Terminal-Bench 2.1, ReactBench, and ADE Bench — run with a 512,000-token cap, temperature 1, top-p 0.95, maximum reasoning enabled, and the Terminus 2 agent harness. On those suites Particle-1 lands at 72.57% on DeepSWE 1.1 (versus 71.15% for DeepSeek V4.1 Flash, i.e. a small gain rather than mere parity), 79.93% on Terminal-Bench 2.1 (versus 81.87%), 64.45% on ADE Bench (versus 67.97%), and 31.46% on ReactBench (versus 33.65%). Particle AI characterizes pass@1 results as falling within the base model's confidence intervals — an accuracy-parity claim, not a uniform win. The trade is explicit and favorable on cost: mean output tokens per task drop from 83,246 to 59,580, a 28.4%–52.2% reduction across the four suites, which translates to a 34.4%–51.6% reduction in average cost per task. GLM 5.3 Flash, evaluated on the same harness, used 98,155 mean output tokens at a 66.82% task pass rate. Serving telemetry from OpenRouter comparison pages shows median generation rates between 77 and 118 tokens per second depending on the paired comparison and routing window (77 tps specifically reported against DeepSeek V4.1 Flash, which measured 97 tps). No first-party Particle AI TTFT, p50/p95 latency, or hardware benchmark exists; all throughput figures are router-observed medians and vary by provider, sampling configuration, and workload shape. Third-party aggregators list divergent pricing ($0.2125–$0.25 input / $0.85–$1.00 output per 1M via resellers) against the $0.05/$0.50 direct listing, which matters for cost modeling. Competitive positioning as of October 2026: Particle-1 does not attempt to out-reason GPT-6 Astra, Claude Opus 5, Gemini 3.8 Pro, Grok 4.6, or Kimi K3. It competes on cost-per-completed-task within the Flash-tier agentic-coding bracket, where its differentiated claim is token discipline — deliberately shorter deliberation loops and leaner tool-call payloads — rather than peak capability. There are no published Particle-1 scores for SWE-bench Verified, SWE-bench Pro, Terminal-Bench 3.0/4.0, HLE (with or without tools), GPQA Diamond, AIME 2026, LiveCodeBench, TAU-2/TAU-3, OSWorld Verified, or any legal/finance/health/cyber suite; none should be inferred from the DeepSeek V4.1 Flash base.
Particle-1 is not an independently architected model — it is DeepSeek V4.1 Flash plus a rank-64 LoRA post-train adapter. The inherited backbone is reported as an MoE transformer with 552B total backbone parameters, 8B active parameters per prompt/prefill token and 16B active per decoded token, DeepSeek Sparse Attention (DSA) with a top-k token indexer (an MLA-lineage latent-attention design), Engram conditional memory, and a causal encoder-decoder formulation. Because only an adapter is trained, tokenizer, embedding space, attention stack, KV-cache geometry, and the 1,048,576-token context / 131,072-token max-output limits carry over unchanged and are confirmed in the Particle-1 OpenRouter listing. Test-time reasoning is enabled (extended thinking tokens) but aggressively compressed: the post-train's explicit objective is fewer reasoning and tool-call tokens per solved task, which is why cost-per-task falls 34.4%–51.6% while pass rates stay inside the base model's confidence intervals. Evaluation used a 512,000-token cap with temperature 1, top-p 0.95, maximum reasoning, and the Terminus 2 agent. Caveat: no first-party Particle AI technical report, parameter count, or attention specification exists; architecture details are attributed to the declared base model and to the adapter-based method, not to an independent Particle-1 spec sheet.
Recommended Workloads & Primary Use Cases
- •Structured outputs: schema enforcement is engine-reduced (json_object supported, strict json_schema not guaranteed)
- •Domain specialization: post-training RL is heavily optimized for software engineering rather than creative prose or open-ended chat
- •Base model boundary: reasoning ceiling tracks DeepSeek V4.1 Flash; will not match GPT-6 Astra on formal mathematics or binary reverse engineering
- •Tool-calling syntax: requires qualified OpenAI tool calling profile; unstructured tool responses may degrade efficiency gains
- •Measured accuracy is parity-with-regression, not improvement across the board: versus DeepSeek V4.1 Flash it loses 1.94 pts on Terminal-Bench 2.1, 3.52 pts on ADE Bench, and 2.19 pts on ReactBench, while gaining ~1.4 pts on DeepSWE 1.1. Teams must validate on their own task distribution before substituting it for the base model.
- •The token savings are conditional on task familiarity. The 28.4%–52.2% reduction comes from training the model to terminate deliberation earlier; on genuinely novel APIs, ambiguous specs, or debugging that requires deep backtracking, the compressed reasoning budget can produce premature 'done' claims or shallow root-cause fixes.
- •Published benchmarks were run under a 512,000-token cap with Terminus 2, temperature 1, top-p 0.95, and maximum reasoning — roughly half the advertised 1,048,576-token window. Behavior at the full 1M context (prefill cost, DSA index recall, and reasoning-quality degradation on very long histories) is uncharacterized publicly.
- •Throughput is router-dependent and inconsistent: OpenRouter comparison pages report 77, 89, 100, 113, and 118 tps medians for the same model. Long agent runs are prefill-bound rather than decode-bound, so TTFT on 100k–500k-token conversation histories dominates wall-clock time regardless of the higher decode rate.
- •It inherits the DeepSeek tool-call schema, chat template, and tokenizer. Deviating from the base prompt format — custom function-calling wrappers, non-DeepSeek tool-result delimiters, or heavy XML/JSON scaffolding — can silently erase the token-efficiency gain and reintroduce verbose reasoning.
- •No published coverage for SWE-bench Verified/Pro, Terminal-Bench 3.0/4.0, HLE (with or without tools), GPQA Diamond, AIME 2026, LiveCodeBench, OSWorld Verified, TAU-2/TAU-3, or any legal/finance/healthcare/cyber suite — do not assume frontier-breadth reasoning from the coding results, and do not import DeepSeek V4.1 Flash scores as proxies.
- •Reseller pricing diverges sharply from the direct listing ($0.05/$0.50 vs $0.2125–$0.25 input and $0.85–$1.00 output per 1M on third-party routers), so realized cost per task can be 4–5x the headline rate depending on route.
Calls to Particle-1 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Particle AI.