[kingsleydon]

Phala provides OpenAI-compatible Chat Completions at https://inference.phala.com/v1 using PHALA_AI_API_KEY; Kev 4B uses POST /v1/systemone. All 21 entries are cataloged as TEE-backed; /v1/aci/attestation attests the gateway, not each inference response.

Entries

20 Chat Completions models and one v2-only System One model use base_model with host overrides; Kev 4B adds a complete shared lab definition. Costs are USD/MTok.

Model ID	base_model	Cost in / out / cache read	Reasoning	Limit deltas vs lab
deepseek/deepseek-v3.2	deepseek/deepseek-v3.2	1 / 1 / 0.5	toggle	context 163,840 (lab 128,000)
deepseek/deepseek-v4-flash-0731	deepseek/deepseek-v4-flash-0731	0.44 / 1.32 / 0.028	toggle + low/high/max	context 1,048,576 (lab 1,000,000); output 393,216 (lab 384,000)
deepseek/deepseek-v4.1-flash-1017	deepseek/deepseek-v4.1-flash-1017	0.75 / 1.6 / 0.15	toggle + low/high/max	context 1,048,576 (lab 1,000,000); output 393,216 (lab 384,000)
google/gemma-4-31b-it	google/gemma-4-31b-it	0.15 / 0.46 / 0.075	toggle	—
jaredpalmer/kev-4b	jaredpalmer/kev-4b	0.042 / 0 / —	— (decision)	context 8,192 (lab 65,536)
meta-llama/llama-3.3-70b-instruct	meta/llama-3.3-70b-instruct	2 / 2 / —	—	context 131,072 (lab 128,000); output 16,384 (lab 4,096)
moonshotai/kimi-k2.6	moonshotai/kimi-k2.6	1.09 / 4.6 / 0.37	toggle	—
moonshotai/kimi-k3	moonshotai/kimi-k3	3 / 15 / 0.3	toggle + low/high/max	—
openai/gpt-oss-120b	openai/gpt-oss-120b	0.15 / 0.6 / —	low/medium/high	—
qwen/qwen-2.5-7b-instruct	alibaba/qwen2.5-7b-instruct	0.1 / 0.2 / —	—	context 32,768 (lab 131,072)
qwen/qwen3-32b	alibaba/qwen3-32b	0.12 / 0.5 / 0.052	toggle + budget	context 40,960 (lab 131,072)
qwen/qwen3-vl-30b-a3b-instruct	alibaba/qwen3-vl-30b-a3b-instruct	0.2 / 0.7 / 0.03	—	context 128,000 (lab 262,144)
qwen/qwen3.5-27b	alibaba/qwen3.5-27b	0.3 / 2.4 / 0.03	toggle	—
qwen/qwen3.5-397b-a17b	alibaba/qwen3.5-397b-a17b	0.55 / 3.5 / 0.225	toggle	—
qwen/qwen3.6-27b	alibaba/qwen3.6-27b	0.32 / 3.25 / 0.03	supported; no controls	—
qwen/qwen3.6-35b-a3b	alibaba/qwen3.6-35b-a3b	0.2 / 1.27 / 0.056	toggle	—
qwen/qwen3.8-27b	alibaba/qwen3.8-27b	0.2 / 2.5 / 0.05	toggle + low/medium/xhigh + budget; default xhigh	context 1,000,000 (lab 262,144); output 262,144 (lab 32,768)
z-ai/glm-5.1	zhipuai/glm-5.1	1.21 / 4.2 / 0.6	supported; no controls	context 202,752 (lab 200,000); output 128,000 (lab 131,072)
z-ai/glm-5.2	zhipuai/glm-5.2	1.26 / 3 / 0.22	high/max	context 1,048,576 (lab 1,000,000)
z-ai/glm-5.3-flash	zhipuai/glm-5.3-flash	0.15 / 0.5 / 0.03	low/high/max	context 1,048,576 (lab 1,000,000)
z-ai/glm-5.3	zhipuai/glm-5.3	1.4 / 4.4 / 0.26	low/high/max	context 1,048,576 (lab 1,000,000)

Reasoning controls

Controls follow native provider entries and relay peers. The leading model comments record the host probes used for deviations (Chat Completions, max_tokens=256, 2026-10-25); these bounded probes do not establish behavior for every prompt.

• reasoning.enabled=false and reasoning_effort=none still returned non-empty reasoning_content on all 3 prompts for GLM 5.1 and Qwen3.6 27B (256 reasoning tokens each). GLM 5.2 off returned 258/256/257; none returned 256/261/258. The off requests still reasoned in those probes. GLM 5.1 has no native graded effort and declares support only.
• reasoning_effort: native/peer sets are shown in the table; model headers record acceptance of DeepSeek V4 low/high/max. GLM 5.3/Flash use effort only, matching their native provider entries. The gateway source confirms reasoning.effort as an alias; the alias was not separately live-tested.
• reasoning.max_tokens: the Qwen3.8 header records an accepted 64-token budget producing 2 reasoning tokens, while no-effort/xhigh both produced 193. The gateway rejects 0; other numeric bounds are unverified and left unspecified. Qwen3 32B keeps the native baseline; its recorded effect test was inconclusive (about 1 token even without a budget).
• Budgets 16 and 128 produced ~256 reasoning tokens on all 3 recorded prompts for Qwen3.5 27B/397B and Qwen3.6 35B, matching the no-budget probes; budget controls remain absent. Qwen3.6 27B requests with reasoning.max_tokens=64 were rejected with HTTP 400 on 3/3 prompts.

[api.chat-completions] compatibility = { reasoning_field = "reasoning_content" } follows Chutes peers on 16 reasoning models. Live checks on 2026-10-25 (max_tokens=256, 10/10 requests HTTP 200):

• DeepSeek V3.2 and Qwen3.8 returned non-empty reasoning_content (117/3 characters); GLM 5.3 and Kimi K2.6 forced tool-call messages included empty reasoning_content/reasoning fields, so those two probes establish field presence and replay acceptance only.
• All four assistant messages were replayed with their reasoning_content, tool calls and tool results: 4/4 follow-ups accepted, including 2/2 non-empty reasoning replays.
• GPT-OSS returned non-empty reasoning and no reasoning_content in 2/2 requests. Its compatibility field stays unset: v1 interleaved.field permits only reasoning_content/reasoning_details (same limitation as the Nebius DeepSeek V4.1 peer). GLM 5.1 and Qwen3.6 27B also returned non-empty reasoning_content in the off-control probes above. The other 10 reasoning models follow peers; their response field was not individually probed.

Catalog input_modalities is text,image for Kimi K2.6/Qwen3 VL and text,image,video for GLM 5.3 Flash. MIME types follow the lab: GLM 5.3 Flash keeps video/mp4, video/x-matroska and video/quicktime; Kimi K2.6/Qwen3 VL remain text/image. None lists PDF, so PDF is omitted. Qwen3 VL's catalog output 32,768 matches its inherited lab value. Context-equal output caps inherit the lab; Qwen2.5 output is 8,192.

Sources

• Catalog: https://inference.phala.com/v1/models (checked 2026-10-25): all 21 selected IDs, is_tee, pricing.prompt/completion (USD/token), pricing.input_cache_read, context_length/max_output_length, input_modalities and supported_parameters; exact MIME types follow the lab files.
• API-key name: Phala's official E2EE guide, Verify Attestation and Verify Signature pages all use PHALA_AI_API_KEY. Each URL returned HTTP 200 and contained that string on 2026-10-25; their .mdx sources on Phala-Network/phala-docs main were also checked.
• Gateway attestation: https://docs.phala.com/phala-cloud/confidential-ai/confidential-model/api-reference/attestation; official logo: https://phala.com/home/logo.svg (both checked 2026-10-25).
• Gateway reasoning wire: https://github.com/Dstack-TEE/private-ai-gateway/blob/main/src/middleware/reasoning.rs; response handling: https://github.com/Dstack-TEE/private-ai-gateway/blob/main/src/middleware/response_transform.rs (checked 2026-10-25; accepts the effort alias and preserves both reasoning and reasoning_content).
• Native/peer reasoning controls: providers/deepseek/models/deepseek-v4-flash.toml, providers/alibaba/models/qwen*.toml, providers/zhipuai/models/glm-5*.toml, providers/moonshotai/models/kimi-*.toml; Kimi K3 relay: providers/nearai/models/moonshotai/kimi-k3.toml; GPT-OSS peer: providers/tinfoil/models/gpt-oss-120b.toml.
• Reasoning compatibility: Chutes providers/chutes/models/deepseek-ai/DeepSeek-V3.2-TEE.toml and providers/chutes/models/moonshotai/Kimi-K2.6-TEE.toml; v1 field limitation: packages/core/src/schema.ts and providers/nebius/models/deepseek-ai/DeepSeek-V4.1-Flash-1017.toml. The response-field/tool-replay counts above were checked live on 2026-10-25.
• Kev: https://huggingface.co/jaredpalmer/kev-4b/raw/main/README.md and https://openrouter.ai/api/v1/models/jaredpalmer/kev-4b/endpoints (checked 2026-10-25). The model card documents the 2026-09-24 release, Apache-2.0 Qwen3.5 adapter/head, states up to 65,536 tokens (validated at 8,192), and typed probabilities without generated text. OpenRouter corroborates System One, text input and Phala's 8,192 context/pricing. The SDK pair follows providers/baseten/models/decision-models/baseten-mammoth-27b-latest.toml; the Kev integration was not live-tested.

Notes for reviewers

• [modalities] in kimi-k2.6, qwen3-vl-30b-a3b-instruct and glm-5.3-flash narrows input to the host catalog. output = ["text"] is required because ModalitiesV2 requires both arrays and modelV2ToV1 validates provider TOML before base_model inheritance (packages/core/src/schema-v2.ts, packages/core/src/translate.ts).

Validation

bun validate and git diff --check pass. bun run v1:check was not run. An equivalent in-place comparison used the repository's generateCatalog, filterCatalogByModelType and MODEL_TYPES, matching the generation in packages/core/script/v1-check.ts, to compare all 22 v1 views against dev 7cd04230c.

The baseline was generated with the two newly added catalog directories temporarily held outside the catalog and restored in finally; all other tracked files matched dev. Only api, api-all and api-chat change, adding Phala and its 20 Chat Completions models. Existing providers, models and fields are unchanged. Phala's Kev lab/provider entries are v2-only and absent from v1. No additional Git worktree was created.

The two uncensored variants (phala/gemma-4-26b-a4b-uncensored and phala/qwen3.8-27b-uncensored) are excluded; Responses/Messages are deferred.

🤖 Generated with Claude Code

[github-actions]

Action items

• [high] [violation] providers/phala/models/qwen/qwen3.8-27b.toml:13 - Check: Relay reasoning controls must copy lab + same-surface peer effort set, not omit graded effort from uncertainty. Why: File claims toggle only while its own header says catalog lists reasoning / reasoning_effort, native alibaba/qwen3.8-27b exposes toggle + low|medium|xhigh + budget, and openrouter/qwen3.8-27b exposes none|low|medium|xhigh; same omission pattern appears in glm-5.1, glm-5.2 and kimi-k2.6 which also cite reasoning_effort then omit effort. Action: Verify live/docs whether Phala forwards reasoning.effort/reasoning_effort for these routes and add peer-supported effort or cite host proof that only reasoning.enabled exists.
• [high] [possible mistake] providers/phala/models/google/gemma-4-31b-it.toml:16 - Check: limit.output must be the host-served max. Why: Lab is 32_768 and chutes/google/gemma-4-31B-turbo-TEE is 65_536; PR claims 262_144 (8x lab, 4x peer) with no peer/catalog support, likely context/output confusion that inflates v1 limit.output. Action: Verify against catalog max_output_tokens/max_completion_tokens and correct or cite exact field.
• [high] [possible mistake] providers/phala/models/qwen/qwen3.6-27b.toml:15 - Check: limit.output accuracy. Why: Lab is 65_536; PR claims 262_140 (off-by-4 typo for 262_144, 4x lab); same 262_144 appears in qwen3.5-397b-a17b, qwen3.6-35b-a3b vs labs 65_536 and OpenRouter 235_929, and qwen3.8-27b claims 262_144 vs lab 32_768. Action: Fix typo and verify all four Qwen outputs against catalog, correcting to verified maxima.
• [high] [possible mistake] providers/phala/models/z-ai/glm-5.3.toml:23 - Check: limit.output must be host-served max. Why: Lab is 131_072 and all peers (tinfoil/glm-5-3, privatemode-ai/glm-5.3, nearai/z-ai/glm-5.3) inherit 131_072 with no output override; PR claims 1_048_576 equalling context (8x). Action: Verify catalog output cap and correct to verified value.
• [medium] [possible mistake] providers/phala/models/moonshotai/kimi-k3.toml:16 - Check: Sampling overrides must be real host deltas. Why: Lab fixes temperature/top_p/top_k=false, native moonshotai/kimi-k3 keeps them fixed and chutes/moonshotai/Kimi-K3-TEE enables only temperature=true; PR enables all three true with only catalog supported_sampling_parameters as evidence and no live test. Action: Verify Phala accepts top_p/top_k on this route or drop to peer-supported set.
• [medium] [possible mistake] providers/phala/models/deepseek/deepseek-v4.1-flash-1017.toml:15 - Check: Sampling deltas must match host API. Why: Native deepseek/deepseek-flash sets only top_k=false (keeps top_p=true); PR sets both top_p=false and top_k=false with header admitting not listed in catalog supported_parameters; not live-tested. Action: Verify docs/live that Phala rejects top_p or inherit top_p from lab.
• [medium] [possible mistake] providers/phala/models/z-ai/glm-5.3-flash.toml:13 - Check: Modalities and effort must match host surface. Why: Lab includes application/pdf, privatemode-ai/glm-5.3-flash narrows to text+image/jpeg|png and nearai/z-ai/glm-5.3-flash to text+2 images; PR keeps video/* but drops pdf unexplained, and claims effort low|high|max while its header says catalog lists only reasoning (no reasoning_effort) and not live-tested. Action: Verify image/video/pdf support and reasoning_effort forwarding on Phala and align modalities/effort or cite.
• [low] [possible mistake] providers/phala/models/deepseek/deepseek-v3.2.toml:16 - Check: Limits and cache.implicit need host evidence. Why: PR claims context 163_840 vs lab 128_000 and Chutes peer 131_072, llama-3.3-70b claims output 16_384 vs lab 4_096 with no peer support, and 14/20 files claim cache={implicit=true} with no caching docs cited, only the models catalog. Action: Cite catalog/docs fields for the divergent limits and caching behavior or correct to verified values.

[github-actions]

Action items

• [high] [possible mistake] providers/phala/models/qwen/qwen3.6-27b.toml:12 - Check: Served output limit vs lab/peer values. Why: output = 262_140 is 4x the lab 65_536 (models/alibaba/qwen3.6-27b.toml) and off-by-4 from the 262_144 used for every other Phala full-context output, indicating a transcription typo rather than a measured host cap. Action: Verify against https://inference.phala.com/v1/models and fix to the catalog's max_output_tokens or remove if it inherits the lab value.
• [high] [possible mistake] providers/phala/models/qwen/qwen3.8-27b.toml:13 - Check: Served context/output vs lab and native controls. Why: context = 1_000_000 / output = 262_144 vs lab 262_144 / 32_768 (models/alibaba/qwen3.8-27b.toml); Alibaba native (providers/alibaba/models/qwen3.8-27b.toml) keeps 262_144 context and Chutes/NEAR peers keep lab context with 65_536 or no output override — a 4x context inflation plus 8x output with output-equals-context pattern matches a context/output confusion. Action: Confirm Phala's actual context_length and max_output_tokens and correct both fields.
• [high] [possible mistake] providers/phala/models/google/gemma-4-31b-it.toml:13 - Check: Served output limit vs lab. Why: output = 262_144 equals the full lab context (models/google/gemma-4-31b-it.toml: context 262_144 / output 32_768) while the cited relay baseline (providers/chutes/models/google/gemma-4-31B-turbo-TEE.toml) serves 65_536; full-context output is implausible without host proof. Action: Verify Phala's output cap and correct to the catalog value or inherit the lab value.
• [high] [possible mistake] providers/phala/models/z-ai/glm-5.3.toml:14 - Check: Served output limit vs lab/native. Why: output = 1_048_576 vs lab 131_072 (models/zhipuai/glm-5.3.toml) and Zhipu native with no output override (providers/zhipuai/models/glm-5.3.toml); same output-equals-context pattern appears in qwen3.5-397b-a17b.toml (262_144 vs lab 65_536) and qwen3.6-35b-a3b.toml (262_144 vs lab 65_536) where Chutes/NEAR/Alibaba peers keep 65_536 or small caps. Action: Verify Phala's max_output_tokens for these three routes and correct to measured caps.
• [medium] [possible mistake] providers/phala/models/deepseek/deepseek-v4.1-flash-1017.toml:9 - Check: Override-only sampling deltas. Why: top_p = false contradicts the lab (models/deepseek/deepseek-v4.1-flash-1017.toml: top_p true) and DeepSeek native (providers/deepseek/models/deepseek-v4-flash.toml sets only top_k = false, keeping top_p); no Phala doc in the diff proves Phala rejects top_p. Action: Remove top_p = false unless Phala docs/live rejection prove it, keeping only top_k = false.
• [medium] [possible mistake] providers/phala/models/moonshotai/kimi-k3.toml:11 - Check: Override-only sampling deltas. Why: temperature/top_p/top_k = true flips lab false/false/false (models/moonshotai/kimi-k3.toml) where Moonshot native (providers/moonshotai/models/kimi-k3.toml) fixes temperature/top_p and NEAR relay peer keeps lab false; the only support cited is generic supported_sampling_parameters with not live-tested. Action: Verify Phala actually accepts non-default sampling on this route or remove the three overrides.
• [medium] [possible mistake] providers/phala/models/deepseek/deepseek-v3.2.toml:2 - Check: Toggle wire path accuracy. Why: Every Phala toggle claims reasoning.enabled = true|false while all cited baselines use different paths (chat_template_kwargs.thinking, enable_thinking, thinking.type = enabled|disabled per Chutes/NEAR/DeepSeek files) and every file states not live-tested, so the exact wire comment required for toggle is unverified and contradicts peers. Action: Confirm Phala's actual on/off field from docs or a live request and correct all toggle headers, or drop toggle if no off switch exists.
• [medium] [violation] providers/phala/models/qwen/qwen3.8-27b.toml:8 - Check: Relay reasoning baseline vs native/peer option set. Why: Alibaba native exposes toggle + effort [low,medium,xhigh] + budget (providers/alibaba/models/qwen3.8-27b.toml); Phala authors toggle alone while stating not live-tested, and the same omission recurs for qwen3-32b / qwen3.5-27b / qwen3.5-397b-a17b / qwen3.6-27b / qwen3.6-35b-a3b where natives have toggle + budget. Leaving controls out from uncertainty violates the relay rule to copy lab/peer controls. Action: Verify whether Phala forwards thinking_budget / reasoning_effort and add budget/effort or cite host proof that the route is toggle-only.
• [low] [possible mistake] providers/phala/models/deepseek/deepseek-v3.2.toml:1 - Check: Source citation mapped to material claims. Why: All 20 files cite only the generic https://inference.phala.com/v1/models for divergent prices, limits, dropped video/PDF modalities, and cache = { implicit = true } (claimed on 17 routes, omitted on 3 with no explanation); a URL alone does not prove the disputed 4-8x limit changes or caching. Action: State in the PR body what that endpoint supports per model (price/limit/modality/cache) or add per-model doc links for the outliers above.

[github-actions]

Action items

• [high] [violation] providers/phala/models/deepseek/deepseek-v4-flash-0731.toml:9 - Check: Relay must copy native effort set for DeepSeek V4. Why: Phala is a multi-model relay; lab baseline is toggle + low/high/max (providers/deepseek/models/deepseek-v4-flash.toml:34, providers/alibaba/models/deepseek-v4-flash-0731.toml:21). File claims only toggle, dropping effort while the same gateway forwards reasoning_effort on other models. Action: Add effort = ["low", "high", "max"] with wire comment or prove with docs/live output that Phala rejects effort for this ID.
• [high] [violation] providers/phala/models/deepseek/deepseek-v4.1-flash-1017.toml:9 - Check: Relay must copy native effort set for DeepSeek V4. Why: Same baseline as above; file is toggle-only with no effort test noted, contradicting lab/peer low/high/max. Action: Add the low/high/max effort list or provide this-host rejection evidence.
• [high] [violation] providers/phala/models/z-ai/glm-5.1.toml:10 - Check: Do not invent effort beyond lab/peers. Why: Lab baseline is toggle-only (providers/zai/models/glm-5.1.toml:19, providers/openrouter/models/z-ai/glm-5.1.toml:11). File drops toggle and invents effort = ["low", "medium", "high"] with no matching native/peer set. Action: Replace with toggle = true unless live meaningful-effect logs for low/medium/high on this host are provided.
• [high] [violation] providers/phala/models/z-ai/glm-5.3-flash.toml:14 - Check: Relay must not drop native effort controls from uncertainty. Why: Baselines are effort = ["low", "high", "max"] with no toggle (providers/zai/models/glm-5.3-flash.toml:18, providers/openrouter/models/z-ai/glm-5.3-flash.toml:16). File keeps only toggle, omitting the graded levels with no effort test. Action: Add effort = ["low", "high", "max"] or prove Phala rejects reasoning_effort for this ID.
• [high] [violation] providers/phala/models/z-ai/glm-5.2.toml:10 - Check: Extra levels need live meaningful effect; do not publish inconclusive controls. Why: File claims effort = ["high", "max"] but its own header says effect comparison was inconclusive due to the 1024-token cap, and drops lab toggle (providers/zai/models/glm-5.2.toml:21). Action: Re-test with a cap that separates levels and confirm reasoning.enabled=false is truly ignored on this host, or align to toggle = true, effort = ["high", "max"].
• [high] [violation] providers/phala/models/qwen/qwen3.8-27b.toml:10 - Check: Baseline effort is native/peer set, not invented high/max. Why: Alibaba baseline is toggle + low/medium/xhigh + budget (providers/alibaba/models/qwen3.8-27b.toml:16); OpenRouter peer is none/low/medium/xhigh (providers/openrouter/models/qwen/qwen3.8-27b.toml:12). File invents low/medium/high/max, drops xhigh/none. Action: Correct to low/medium/xhigh (+ toggle only if reasoning.enabled=false verified) or show per-level reasoning-token separation proving high/max are distinct host levels.
• [high] [possible mistake] providers/phala/models/qwen/qwen3.6-27b.toml:8 - Check: No controls means affirmatively always-on, not untested. Why: Both baselines expose a toggle (providers/alibaba/models/qwen3.6-27b.toml:21, providers/openrouter/models/qwen/qwen3.6-27b.toml:11), and every sibling Phala Qwen file on the same reasoning.enabled gateway claims the toggle works; single-model “off ignored” contradicts the shared wire path. Action: Verify reasoning.enabled=false truly still reasons on this ID and that reasoning_effort/thinking_budget are also rejected, or add toggle = true.
• [medium] [possible mistake] providers/phala/models/qwen/qwen-2.5-7b-instruct.toml:7 - Check: Provider file must include real limit deltas; PR body must match file. Why: Body table advertises output 32,768 vs lab 8,192, but file sets only context = 32_768 and inherits lab output 8,192. Action: Set [limit] output = 32_768 if catalog max_output_length is authoritatively 32,768, or correct the body table if the catalog value echoes context and lab 8,192 was intentionally kept.

[github-actions]

Action items

• [high] [violation] providers/phala/models/z-ai/glm-5.1.toml:17 - Check: Baseline effort = native/peer set, never invent low/medium/high. Why: Lab zhipuai/glm-5.1 is reasoning, first-party providers/zai/models/glm-5.1.toml is toggle only, providers/tensorx/models/z-ai/glm-5.1.toml is toggle only, providers/alibaba/models/glm-5.1.toml is toggle+budget; no source has low/medium/high for 5.1, yet Phala authors effort = ["low", "medium", "high"] and drops toggle. Action: Remove invented effort and restore toggle = true unless Phala docs prove a gateway effort mapping with meaningful effect beyond “accepted”; keep the top-of-file wire comment.
• [high] [violation] providers/phala/models/z-ai/glm-5.2.toml:17 - Check: Copy underlying controls from lab + same-surface peers on relays. Why: providers/zai/models/glm-5.2.toml, providers/alibaba/models/glm-5.2.toml, and providers/tensorx/models/z-ai/glm-5.2.toml are all toggle = true + effort = ["high", "max"]; Phala drops toggle on a single “off ignored, 258 tokens” probe. Action: Restore toggle = true or provide stronger evidence the gateway does not forward reasoning.enabled/thinking.type for this ID (multiple prompts, gateway docs), not one sample.
• [high] [possible mistake] providers/phala/models/qwen/qwen3.6-27b.toml:17 - Check: Do not leave controls out from uncertainty on relays; reasoning = { supported = true } means affirmatively always-on. Why: Native providers/alibaba/models/qwen3.6-27b.toml is toggle+budget, and same-gateway siblings qwen3.5-27b/qwen3.6-35b-a3b in this PR claim toggle works via the same src/middleware/reasoning.rs; a single “off returned 256/256 tokens, max_tokens 400” is inconsistent with the shared middleware. Action: Re-test reasoning.enabled=false/reasoning_effort=none and reasoning.max_tokens; author toggle (plus budget if forwarded like qwen3-32b) or prove model-specific always-on with docs.
• [high] [possible mistake] providers/phala/models/qwen/qwen3.8-27b.toml:464 - Check: Limits must be host-served deltas; never derive budget min/max from limit.output/context. Why: Lab is context 262_144/output 32_768, native providers/alibaba/models/qwen3.8-27b.toml inherits it, peers are smaller (providers/groq/models/qwen/qwen3.8-27b.toml 131_072/16_384); Phala claims 1_000_000/262_144 (4x/8x lab) while budget = { min = 1, max = 262_144 } exactly equals its own output and differs from native min = 0. Action: Verify catalog context_length/max_output_length against https://inference.phala.com/v1/models, confirm budget bounds come from the gateway reasoning.max_tokens range/test not the output cap, and fix min or document why 0 is rejected.
• [medium] [possible mistake] providers/phala/models/moonshotai/kimi-k2.6.toml:219 - Check: Provider base_model files are override-only with real deltas. Why: Lab input includes 8 image types + 9 video types; Phala narrows to image-only with no catalog/test note, same pattern as qwen/qwen3-vl-30b-a3b-instruct.toml:340 (drops all video vs lab) and z-ai/glm-5.3-flash.toml:549 (drops video + application/pdf vs lab). Action: Verify each ID’s catalog modalities or live video/PDF probe and cite it in the header, or delete the [modalities] override to inherit the lab.
• [medium] [possible mistake] providers/phala/models/qwen/qwen3-vl-30b-a3b-instruct.toml:347 - Check: Do not restate identical limits; author only real deltas. Why: File sets context = 128_000 but omits output, inheriting lab 32_768 which happens to equal the header’s catalog max_output_length 32768 — correct if verified, but header cites only context_length; if catalog output differs the inherit understates/overstates the cap. Action: Confirm catalog max_output_length and either keep inheritance with an explicit header value or author explicit output.
• [low] [possible mistake] providers/phala/models/qwen/qwen3.5-27b.toml:364 - Check: budget only for a real reasoning-budget field on this path. Why: Same gateway claims reasoning.max_tokens exists (qwen3-32b, qwen3.8-27b author budget), native providers/alibaba/models/qwen3.5-27b.toml and qwen3.5-397b-a17b/qwen3.6-35b-a3b peers are toggle+budget, yet Phala omits budget on qwen3.5-27b, qwen3.5-397b-a17b, and qwen3.6-35b-a3b from one 64 → ~256 tokens sample. Action: Re-test budget scaling with multiple values or author budget = {} if the field is accepted, consistent with qwen3-32b; same for the two sibling files.

[github-actions]

Action items

• [high] [possible mistake] providers/phala/models/qwen/qwen3.6-27b.toml:8 - Check: Reasoning budget must be authored when the host exposes a real reasoning-token budget field. Why: The file claims always-on (reasoning = { supported = true }) but its own header reports reasoning.max_tokens returned 400 on 3/3 vs 256 tokens for the no-budget control, which indicates the budget field had an effect and contradicts no-caller-control. Action: Re-verify whether reasoning.max_tokens controls reasoning tokens on this host; if it does, add budget = {} (with verified min/max only) matching the alibaba/qwen3.6-27b native baseline, or clarify the comment if 400 was not reasoning tokens.
• [medium] [possible mistake] providers/phala/models/z-ai/glm-5.3-flash.toml:10 - Check: [modalities] must match the cited catalog capabilities. Why: The header cites catalog input_modalities text,image,video but the file publishes only ["text", "image/jpeg", "image/png"], omitting video entirely because generic video has no exact MIME, which understates the catalog and diverges from the lab's video/mp4, video/x-matroska, video/quicktime. Action: Include the supported video modalities (lab MIME types or generic video if exact mapping is unknown) or provide host evidence that video is not actually served.

[github-actions]

Action items

• [medium] [possible mistake] providers/phala/models/qwen/qwen3.8-27b.toml:10 - Check: Reasoning budget min/max must be verified on this host; use budget = {} when bounds are not verified and never derive bounds from limit.output/context. Why: budget = { min = 1, max = 262_144 } sets max to exactly the file's limit.output = 262_144, but the header sources max from native thinking_budget 0-262144 (providers/alibaba/models/qwen3.8-27b.toml), not a Phala live test, while min = 1 is sourced from gateway code; the host max is unverified and matches the output cap. Action: Verify reasoning.max_tokens upper bound live on Phala for this model or relax to budget = {} (keeping toggle/effort/default as verified).

[github-actions]

No actionable findings.

[github-actions]

Action items

• [high] [violation] providers/phala/models/z-ai/glm-5.3-flash.toml:29 - Check: Reasoning baseline must be lab + same-surface peers, do not invent toggle. Why: Lab models/zhipuai/glm-5.3-flash.toml and first-party providers/zhipuai/models/glm-5.3-flash.toml:10 / providers/zai/models/glm-5.3-flash.toml:8 state thinking.type only accepts enabled; peers providers/novita-ai/models/zai-org/glm-5.3-flash.toml:2 (thinking.type=disabled does not suppress) and providers/openrouter/models/z-ai/glm-5.3-flash.toml:16 are effort-only ["low","high","max"] with no toggle. PR adds toggle = true contradicting all three. Action: Remove toggle, keep effort = ["low","high","max"] unless you provide gateway docs plus full request/response proving reasoning.enabled=false yields zero reasoning_content where Novita proves it does not.
• [high] [violation] providers/phala/models/z-ai/glm-5.3.toml:22 - Check: Do not invent toggle beyond lab/peers. Why: Lab models/zhipuai/glm-5.3.toml and providers/zhipuai/models/glm-5.3.toml:9 are effort-only low|high|max (thinking.type only accepts enabled); peers are providers/novita-ai/models/zai-org/glm-5.3.toml:10 (["none","low","high","max"], no toggle) and providers/openrouter/models/z-ai/glm-5.3.toml:11 (["low","high","max"], no toggle). No lab/peer has toggle. Action: Drop toggle; re-test reasoning.enabled=false vs reasoning_effort=none and author either ["none","low","high","max"] (if none suppresses) or ["low","high","max"] (if always-on with effort only).
• [high] [possible mistake] providers/phala/models/z-ai/glm-5.2.toml:11 - Check: Missing off-control where lab/peers expose it. Why: Native providers/zhipuai/models/glm-5.2.toml:11 is toggle + ["high","max"]; peer providers/novita-ai/models/zai-org/glm-5.2.toml:9 is ["none","high","max"] where none suppresses. PR claims both reasoning.enabled=false and reasoning_effort=none returned ~256 tokens and authors no off (["high","max"] only), directly contradicting Novita on the same model. Action: Re-verify with reasoning_content traces whether either off mechanism suppresses reasoning on Phala; add toggle and/or none to match the verified behavior.
• [high] [possible mistake] providers/phala/models/z-ai/glm-5.1.toml:12 - Check: Missing toggle on relay where lab and peers have it. Why: Native providers/zhipuai/models/glm-5.1.toml:21, providers/novita-ai/models/zai-org/glm-5.1.toml:9, and providers/openrouter/models/z-ai/glm-5.1.toml:11 are all toggle = true. PR authors always-on reasoning = { supported = true } based only on own 3/3 off-ignored probes. Action: Re-test thinking.type=disabled forwarding on Phala and provide evidence; if off cannot be made to work, explain why Phala differs from two relays that honor it, else add toggle = true with wire comment.
• [high] [possible mistake] providers/phala/models/qwen/qwen3.6-27b.toml:11 - Check: Missing controls where lab + peers expose toggle/budget. Why: Native providers/alibaba/models/qwen3.6-27b.toml:21 is toggle + budget, peers providers/novita-ai/models/qwen/qwen3.6-27b.toml:9 (toggle + budget) and providers/openrouter/models/qwen/qwen3.6-27b.toml:11 (toggle) both honor off. PR claims reasoning.enabled=false/none ignored and reasoning.max_tokens rejected and authors no controls, making Phala the sole outlier among 12 sibling Qwen routes on the same gateway that do honor toggle. Action: Re-test toggle/budget forwarding for this ID; add toggle = true (and budget = {} if thinking_budget is accepted) or prove a model-specific gateway exception with traces.
• [medium] [possible mistake] providers/phala/models/qwen/qwen3.8-27b.toml:44 - Check: limit.output must be catalog max_output_length, not thinking_budget max. Why: Lab models/alibaba/qwen3.8-27b.toml:30 is 32_768; peers are 131_072 (Novita) and 235_929 (OpenRouter). PR sets 262_144, which exactly equals native thinking_budget 0-262144 max (providers/alibaba/models/qwen3.8-27b.toml:3). Comment cites catalog max_output_length 262144 but number matches budget cap, suggesting field confusion. Action: Confirm which catalog field yields 262144 and cite it; if it is thinking_budget max, inherit lab output or use peer-verified cap instead.

[github-actions]

Action items

• [high] [violation] providers/phala/models/moonshotai/kimi-k2.6.toml:11 - Check: Override-only provider files after base_model. Why: output = ["text"] is identical to models/moonshotai/kimi-k2.6.toml output; only input is a real delta (drops video). Restating identical arrays violates AGENTS.md. Action: Remove the output line and keep only the narrowed input.
• [high] [violation] providers/phala/models/qwen/qwen3-vl-30b-a3b-instruct.toml:10 - Check: Override-only provider files after base_model. Why: output = ["text"] is identical to models/alibaba/qwen3-vl-30b-a3b-instruct.toml output; only input is a real delta (drops video). Action: Remove the output line and keep only the narrowed input.
• [high] [violation] providers/phala/models/z-ai/glm-5.3-flash.toml:10 - Check: Override-only provider files after base_model. Why: output = ["text"] is identical to models/zhipuai/glm-5.3-flash.toml output; only input is a real delta (drops application/pdf). Action: Remove the output line and keep only the narrowed input.

[github-actions]

No actionable findings.

[kingsleydon]

This adds Phala (21 models, including Kev 4B via System One) and is now marked reviewer: ready. Notes on the schema-required output arrays and the live-tested reasoning-control deviations are in the description. Ready for review. Thank you!

[github-actions]

No actionable findings.