Choose your model

Choosing a model

Start cheap, profile, upgrade when the task gets harder. Switching is a one-string change. The full live list with INR/M pricing is at /models.

Aliases vs long form

Ollima accepts two forms:

Alias — short, branded, marketing-friendly. e.g. deepseek-v4. We resolve it to a provider at request time. Add or rename aliases on the dashboard's Routes page.
Long form — explicit: tensorzero::model_name::<provider>::<upstream-slug>. Bypasses alias resolution. Use when you want a guaranteed specific provider.

Common picks

Alias	What it's for	Approx ₹/1M input
`deepseek-v4-flash`	Cheap + fast. Prototype here, and for autocomplete + classification.	~ ₹5
`deepseek-v4`	Workhorse. Chat, tools, RAG, hard multi-file refactors.	~ ₹50
`qwen3-plus`	Balanced quality — very long context (~1M tokens).	~ ₹30
`qwen3-32b`	Reasoning model. Math + planning.	~ ₹25

Premium coding lane

Strong code models in the Ollima catalog, every one priced under $3 / 1M output — the metric that matters for output-heavy coding. Same OpenAI-shape call, just change model.

Alias	What it's for	Context	₹/1M in · out
`deepseek-v4`	Flagship coder. Hard refactors, multi-file reasoning.	1M	~ ₹37 · ₹74
`deepseek-v4-flash`	Cheapest strong coder. Autocomplete, quick edits.	1M	~ ₹8 · ₹15
`qwen3-plus`	Balanced quality + long context.	1M	~ ₹27 · ₹109
`qwen3-35b`	MoE — fast and cheap for everyday coding.	262K	~ ₹13 · ₹85
`minimax-m3`	Agentic / tool-use coding. 80.5% SWE-bench Verified.	1M	~ ₹26 · ₹102
`step-3-flash`	Fast low-cost generalist coder.	256K	~ ₹17 · ₹98
`mimo-v2`	MiMo — agentic coder, long context. Reasoning model.	1M	~ ₹37 · ₹74

deepseek-v4 (80.6% SWE-bench Verified) and minimax-m3 (80.5%) are the current open-weights benchmark leaders — both in the under-$3 lane above.

mimo-v2 (and qwen3-32b) are reasoning models — give them a generous max_tokens and don't use them for autocomplete, or they spend the whole budget thinking and return empty content (verified). For inline completion use deepseek-v4-flash.

Premium tier (output above $3/1M)

One step up in price for top benchmark quality:

Alias	What it's for	Context	₹/1M in · out
`glm-5.1`	GLM — top-ranked coding/agentic; structured output.	203K	~ ₹83 · ₹262

Further premium models above the $3 bar (kimi-k2.7-code, glm-5.2) aren't in the built-in catalog yet — bring them via BYOK, or ask us to add them. INR figures are approximate (USD list price × an indicative FX); the dashboard bills on the live rate.

Measured latency

Real numbers, measured through the gateway on 2026-06-18 — a short coding prompt capped at 160 output tokens, best of two streamed runs. TTFT is time-to-first-token (what interactive coding feels like); Total is the full 160-token response.

Model	TTFT	Total (160 tok)
`deepseek-v4-flash`	0.05s	0.25s
`deepseek-v4`	0.08s	0.34s
`qwen3-32b`	0.14s	0.44s
`qwen3-plus`	0.46s	0.74s
`glm-5.1`	0.16s	1.12s
`step-3-flash`	0.19s	1.25s
`qwen3-35b`	0.25s	1.49s
`deepseek-v4-flash`	0.55s	1.64s
`mimo-v2`	0.39s	2.09s
`minimax-m3`	0.84s	2.45s
`deepseek-v4`	1.18s	5.53s
`qwen3-plus`	1.20s	~19s ⚠

The pattern: our instant lane — deepseek-v4-flash, qwen3-32b, qwen3-plus — finishes a 160-token reply in under a second. The larger coding models land in 1–2.5s, the flagship deepseek-v4 takes ~5.5s, and qwen3-plus was a high-variance outlier (up to ~19s under load) in this sample. For latency-critical work (autocomplete, fast chat) prefer the instant lane or deepseek-v4-flash. These are single-region, point-in-time samples — treat them as indicative, not an SLA.

Switching models

Change the model string. That's it. The rest of the request stays identical. Watch the Logs page on the dashboard to compare latency + token consumption side-by-side.

Custom aliases & failover

On the dashboard's Routes page, map any short name to one or more Ollima models — the first is preferred, the rest are failover targets:

# Pseudo-config — actual edits happen on the dashboard UI
alias: my-fast-router
targets:
  - model: deepseek-v4       # primary
  - model: deepseek-v4-flash   # falls back here if the primary errors
timeout_ms: 30000
max_retries: 1

Then your client uses model: "my-fast-router". Order = preference; subsequent entries are failover targets.